Libc-825.25.tar.gz

[apple/libc.git] / gen / scalable_malloc.c

diff --git a/gen/scalable_malloc.c b/gen/scalable_malloc.c
index 1a47195239b60bcf261575e62cc5da17308f6bb1..6ac0ae662f6c3c4b704c074943838e398b6d17e6 100644 (file)
--- a/gen/scalable_malloc.c
+++ b/gen/scalable_malloc.c
@@ -1,10 +1,8 @@
 /*
 /*

- * Copyright (c) 1999 Apple Computer, Inc. All rights reserved.
+ * Copyright (c) 1999, 2006 Apple Computer, Inc. All rights reserved.

  *
  * @APPLE_LICENSE_HEADER_START@
  * 
  *
  * @APPLE_LICENSE_HEADER_START@
  * 

- * Copyright (c) 1999-2003 Apple Computer, Inc.  All Rights Reserved.
- * 

  * This file contains Original Code and/or Modifications of Original Code
  * as defined in and that are subject to the Apple Public Source License
  * Version 2.0 (the 'License'). You may not use this file except in
  * This file contains Original Code and/or Modifications of Original Code
  * as defined in and that are subject to the Apple Public Source License
  * Version 2.0 (the 'License'). You may not use this file except in


@@ -25,13 +23,18 @@
 
 /* Author: Bertrand Serlet, August 1999 */
 
 
 /* Author: Bertrand Serlet, August 1999 */
 

-#import "scalable_malloc.h"
+#include "scalable_malloc.h"
+#include "malloc_printf.h"
+#include "_simple.h"

 
 

-#import <pthread_internals.h>
+#include <pthread_internals.h>

 
 

-#import <unistd.h>
-#import <libc.h>
+#include <unistd.h>
+#include <libc.h>

 #include <mach/vm_statistics.h>
 #include <mach/vm_statistics.h>

+#include <mach/mach_init.h>
+#include <sys/types.h>
+#include <sys/mman.h>

 
 /********************* DEFINITIONS     ************************/
 
 
 /********************* DEFINITIONS     ************************/
 


@@ -41,34 +44,43 @@
 
 #if DEBUG_MALLOC
 #warning DEBUG_MALLOC ENABLED
 
 #if DEBUG_MALLOC
 #warning DEBUG_MALLOC ENABLED

-#define INLINE
-#define CHECK_LOCKED(szone, fun)       {       \
-    if (__is_threaded && TRY_LOCK(szone->lock)) {                      \
+# define INLINE
+# define ALWAYSINLINE
+# define CHECK_LOCKED(szone, fun)                                              \
+do {                                                                           \
+    if (__is_threaded && TRY_LOCK(szone->lock)) {                              \

        malloc_printf("*** lock was not set %p in %s\n", szone->lock, fun);     \
        malloc_printf("*** lock was not set %p in %s\n", szone->lock, fun);     \

-    }                                          \
-}
+    }                                                                          \
+} while (0)

 #else
 #else

-#define INLINE __inline__
-#define CHECK_LOCKED(szone, fun)       {}
+# define INLINE        __inline__
+# define ALWAYSINLINE __attribute__((always_inline))
+# define CHECK_LOCKED(szone, fun)      {}

 #endif
 
 #endif
 

-#define PAGE_SIZE_FIXED        1       // flip if the page size becomes variable, one day
-#if PAGE_SIZE_FIXED
+/*
+ * Access to global variables is slow, so optimise our handling of vm_page_size
+ * and vm_page_shift.
+ */
+#define _vm_page_size  vm_page_size    /* to get to the originals */
+#define _vm_page_shift vm_page_shift
+#define vm_page_size   4096            /* our normal working sizes */

 #define vm_page_shift  12
 #define vm_page_shift  12

-#else
-static unsigned vm_page_shift = 0; // guaranteed to be intialized by zone creation
-#endif

 
 typedef unsigned short msize_t; // a size in multiples of SHIFT_SMALL_QUANTUM or SHIFT_TINY_QUANTUM
 
 
 typedef unsigned short msize_t; // a size in multiples of SHIFT_SMALL_QUANTUM or SHIFT_TINY_QUANTUM
 

+typedef union {
+       void       *p;
+       uintptr_t       u;
+} ptr_union;
+

 typedef struct {
 typedef struct {

-    unsigned   checksum;
-    void       *previous;
-    void       *next;
+       ptr_union previous;
+       ptr_union next;

 } free_list_t;
 
 typedef struct {
 } free_list_t;
 
 typedef struct {

-    unsigned   address_and_num_pages;
+    uintptr_t  address_and_num_pages;

     // this type represents both an address and a number of pages
     // the low bits are the number of pages; the high bits are the address
     // note that the exact number of bits used for depends on the page size
     // this type represents both an address and a number of pages
     // the low bits are the number of pages; the high bits are the address
     // note that the exact number of bits used for depends on the page size


@@ -79,35 +91,124 @@ typedef unsigned char      grain_t;
 
 #define CHECK_REGIONS                  (1 << 31)
 
 
 #define CHECK_REGIONS                  (1 << 31)
 

-#define CHECKSUM_MAGIC                 0x357B
-

 #define MAX_RECORDER_BUFFER    256
 
 /********************* DEFINITIONS for tiny    ************************/
 
 #define MAX_RECORDER_BUFFER    256
 
 /********************* DEFINITIONS for tiny    ************************/
 

-#define        SHIFT_TINY_QUANTUM                      4       // Required for AltiVec
+/*
+ * Memory in the Tiny range is allocated from regions (heaps) pointed to by the szone's tiny_regions
+ * pointer.
+ *
+ * Each region is laid out as a heap, followed by a header block, all within
+ * a 1MB (2^20) block.  This means there are 64520 16-byte blocks and the header is
+ * 16138 bytes, making the total 1048458 bytes, leaving 118 bytes unused.
+ * The header block is arranged:
+ *
+ * 0xfc080
+ *     header bits
+ * 0xfe001
+ *     0xffffffff pad word
+ * 0xfe005
+ *     in-use bits
+ * 0xfff86
+ *     pad word (not written)
+ * 0xfff8a-0xfffff
+ *     unused
+ *
+ * Each bitfield comprises NUM_TINY_BLOCKS bits, and refers to the corresponding TINY_QUANTUM block
+ * within the heap.
+ *
+ * The bitfields are used to encode the state of memory within the heap.  The header bit indicates
+ * that the corresponding quantum is the first quantum in a block (either in use or free).  The
+ * in-use bit is set for the header if the block has been handed out (allocated).  If the header
+ * bit is not set, the in-use bit is invalid.
+ *
+ * The szone maintains an array of 32 freelists, each of which is used to hold 
+ * free objects of the corresponding quantum size.
+ *
+ * A free block is laid out depending on its size, in order to fit all free
+ * blocks in 16 bytes, on both 32 and 64 bit platforms.  One quantum blocks do
+ * not store their size in the block, instead relying on the header information
+ * to determine their size.  Blocks of two or more quanta have room to store
+ * their size in the block, and store it both after the 'next' pointer, and in
+ * the last 2 bytes of the block.
+ *
+ * 1-quantum block
+ * Offset (32-bit mode)        (64-bit mode)
+ * 0x0          0x0      : previous
+ * 0x4          0x08     : next
+ * end          end
+ *
+ * >1-quantum block
+ * Offset (32-bit mode)        (64-bit mode)
+ * 0x0          0x0      : previous
+ * 0x4          0x08     : next
+ * 0x8          0x10     : size (in quantum counts)
+ * end - 2      end - 2  : size (in quantum counts)
+ * end          end
+ *
+ * All fields are pointer-sized, except for the size which is an unsigned short.
+ *
+ */
+
+#define SHIFT_TINY_QUANTUM                     4       // Required for AltiVec

 #define        TINY_QUANTUM                            (1 << SHIFT_TINY_QUANTUM)
 
 #define        TINY_QUANTUM                            (1 << SHIFT_TINY_QUANTUM)
 

-#define FOLLOWING_TINY_PTR(ptr,msize)  (((char *)(ptr)) + ((msize) << SHIFT_TINY_QUANTUM))
+#define FOLLOWING_TINY_PTR(ptr,msize)          (((unsigned char *)(ptr)) + ((msize) << SHIFT_TINY_QUANTUM))

 
 

-#define NUM_TINY_SLOTS                 32      // number of slots for free-lists
+#define NUM_TINY_SLOTS                         32      // number of slots for free-lists

 
 

-#define SHIFT_NUM_TINY_BLOCKS  16
-#define NUM_TINY_BLOCKS                (1 << SHIFT_NUM_TINY_BLOCKS)
-#define TINY_BLOCKS_ALIGN      (SHIFT_NUM_TINY_BLOCKS + SHIFT_TINY_QUANTUM)
-#define TINY_REGION_SIZE       ((NUM_TINY_BLOCKS * TINY_QUANTUM + (NUM_TINY_BLOCKS >> 2) + 8 + (1 << vm_page_shift) - 1) & ~ ((1 << vm_page_shift) - 1))       // enough room for the data, followed by the bit arrays (2-bits per block) plus 2 words of padding as our bitmap operators overflow, plus rounding to the nearest page
+#define NUM_TINY_BLOCKS                                64520
+#define SHIFT_TINY_CEIL_BLOCKS                 16 // ceil(log2(NUM_TINY_BLOCKS))
+#define NUM_TINY_CEIL_BLOCKS                   (1 << SHIFT_TINY_CEIL_BLOCKS)
+#define TINY_BLOCKS_ALIGN                      (SHIFT_TINY_CEIL_BLOCKS + SHIFT_TINY_QUANTUM)

 
 

-#define TINY_FREE_SIZE(ptr)    (((msize_t *)(ptr))[6])
-// At the end of free blocks, we stick the size (for enabling coalescing)
-#define TINY_PREVIOUS_MSIZE(ptr)               ((msize_t *)(ptr))[-1]
+/*
+ * Enough room for the data, followed by the bit arrays (2-bits per block) plus 2 words of padding
+ * as our bitmap operators overflow, plus rounding to the nearest page.
+ */
+#define TINY_HEADER_SIZE       ((NUM_TINY_BLOCKS >> 2) + 8)
+#define TINY_REGION_SIZE       ((NUM_TINY_BLOCKS * TINY_QUANTUM + TINY_HEADER_SIZE + vm_page_size - 1) & ~ (vm_page_size - 1))

 
 

+/*
+ * Beginning and end pointers for a region's heap.
+ */
+#define TINY_REGION_ADDRESS(region)            ((void *)(region))
+#define TINY_REGION_END(region)                        (TINY_REGION_ADDRESS(region) + (NUM_TINY_BLOCKS * TINY_QUANTUM))

 
 

-#define TINY_REGION_ADDRESS(region)            ((region) << TINY_BLOCKS_ALIGN)
-#define TINY_REGION_END(region)                (TINY_REGION_ADDRESS(region)+(1 << TINY_BLOCKS_ALIGN))
+/*
+ * Locate the heap base for a pointer known to be within a tiny region.
+ */
+#define TINY_REGION_FOR_PTR(_p)                        ((void *)((uintptr_t)(_p) & ~((1 << TINY_BLOCKS_ALIGN) - 1)))

 
 

-typedef unsigned short tiny_region_t;
+/*
+ * Convert between byte and msize units.
+ */
+#define TINY_BYTES_FOR_MSIZE(_m)               ((_m) << SHIFT_TINY_QUANTUM)
+#define TINY_MSIZE_FOR_BYTES(_b)               ((_b) >> SHIFT_TINY_QUANTUM)
+
+#ifdef __LP64__
+# define TINY_FREE_SIZE(ptr)                   (((msize_t *)(ptr))[8])
+#else
+# define TINY_FREE_SIZE(ptr)                   (((msize_t *)(ptr))[4])
+#endif
+#define TINY_PREVIOUS_MSIZE(ptr)               ((msize_t *)(ptr))[-1]
+
+/*
+ * Locate the block header for a pointer known to be within a tiny region.
+ */
+#define TINY_HEADER_START                      (NUM_TINY_BLOCKS * TINY_QUANTUM)
+#define TINY_BLOCK_HEADER_FOR_PTR(_p)          ((void *)((uintptr_t)TINY_REGION_FOR_PTR(_p) + TINY_HEADER_START))

 
 

-#define INITIAL_NUM_TINY_REGIONS       24      // must be even for szone to be aligned
+/*
+ * Locate the inuse map for a given block header pointer.
+ */  
+#define TINY_INUSE_FOR_HEADER(_h)              ((void *)((uintptr_t)(_h) + (NUM_TINY_BLOCKS >> 3) + 4))
+
+/*
+ * Compute the bitmap index for a pointer known to be within a tiny region.
+ */
+#define TINY_INDEX_FOR_PTR(_p)                         (((uintptr_t)(_p) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_CEIL_BLOCKS - 1))

 
 #define TINY_CACHE     1       // This governs a last-free cache of 1 that bypasses the free-list
 
 
 #define TINY_CACHE     1       // This governs a last-free cache of 1 that bypasses the free-list
 


@@ -117,43 +218,105 @@ typedef unsigned short tiny_region_t;
 
 /********************* DEFINITIONS for small   ************************/
 
 
 /********************* DEFINITIONS for small   ************************/
 

-/* We store the meta bits on the side in two bytes, as follows:
-- high order bit SMALL_IS_FREE is set when the block is avail (and starts here)
-- when block size, expressed in SMALL_QUANTUM, is the other 15 bits
-- else 0 signifies this block is in the middle of another block
-*/
+/*
+ * Memory in the Small range is allocated from regions (heaps) pointed to by the szone's small_regions
+ * pointer.
+ *
+ * Each region is laid out as a heap, followed by the metadata array, all within an 8MB (2^23) block.
+ * The array is arranged as an array of shorts, one for each SMALL_QUANTUM in the heap.
+ * This means there are 16320 512-blocks and the array is 16320*2 bytes, which totals 8388480, leaving
+ * 128 bytes unused.
+ *
+ * The MSB of each short is set for the first quantum in a free block.  The low 15 bits encode the
+ * block size (in SMALL_QUANTUM units), or are zero if the quantum is not the first in a block.
+ *
+ * The szone maintains an array of 32 freelists, each of which is used to hold free objects
+ * of the corresponding quantum size.
+ *
+ * A free block is laid out as:
+ *
+ * Offset (32-bit mode)        (64-bit mode)
+ * 0x0          0x0      : previous
+ * 0x4          0x08     : next
+ * 0x8          0x10     : size (in quantum counts)
+ * end - 2             end - 2  : size (in quantum counts)
+ * end                 end
+ *
+ * All fields are pointer-sized, except for the size which is an unsigned short.
+ *
+ */

 
 

-#define SMALL_IS_FREE  (1 << 15)
+#define SMALL_IS_FREE                          (1 << 15)

 
 #define        SHIFT_SMALL_QUANTUM                     (SHIFT_TINY_QUANTUM + 5)        // 9
 #define        SMALL_QUANTUM                           (1 << SHIFT_SMALL_QUANTUM) // 512 bytes
 
 
 #define        SHIFT_SMALL_QUANTUM                     (SHIFT_TINY_QUANTUM + 5)        // 9
 #define        SMALL_QUANTUM                           (1 << SHIFT_SMALL_QUANTUM) // 512 bytes
 

-#define FOLLOWING_SMALL_PTR(ptr,msize) (((char *)(ptr)) + ((msize) << SHIFT_SMALL_QUANTUM))
+#define FOLLOWING_SMALL_PTR(ptr,msize)         (((unsigned char *)(ptr)) + ((msize) << SHIFT_SMALL_QUANTUM))

 
 

-#define NUM_SMALL_SLOTS                        32      // number of slots for free-lists
+#define NUM_SMALL_SLOTS                                32      // number of slots for free-lists

 
 

-#define SHIFT_NUM_SMALL_BLOCKS         14      // we can only represent up to 1<<15 for msize; but we chose to stay even below that to avoid the convention msize=0 => msize = (1<<15)
-#define NUM_SMALL_BLOCKS               (1 << SHIFT_NUM_SMALL_BLOCKS)
-#define SMALL_BLOCKS_ALIGN             (SHIFT_NUM_SMALL_BLOCKS + SHIFT_SMALL_QUANTUM) // 23
-#define SMALL_REGION_SIZE                      (NUM_SMALL_BLOCKS * SMALL_QUANTUM + NUM_SMALL_BLOCKS * 2)       // data + meta data
+/*
+ * We can only represent up to 1<<15 for msize; but we choose to stay even below that to avoid the
+ * convention msize=0 => msize = (1<<15)
+ */
+#define NUM_SMALL_BLOCKS                       16320
+#define SHIFT_SMALL_CEIL_BLOCKS                        14 // ceil(log2(NUM_SMALL_BLOCKs))
+#define NUM_SMALL_CEIL_BLOCKS                  (1 << SHIFT_SMALL_CEIL_BLOCKS)
+#define SMALL_BLOCKS_ALIGN                     (SHIFT_SMALL_CEIL_BLOCKS + SHIFT_SMALL_QUANTUM) // 23
+#define SMALL_ARRAY_SIZE                       (NUM_SMALL_BLOCKS * 2)
+#define SMALL_REGION_SIZE                      ((NUM_SMALL_BLOCKS * SMALL_QUANTUM + SMALL_ARRAY_SIZE + vm_page_size - 1) & ~ (vm_page_size - 1))       // data + meta data

 
 #define SMALL_PREVIOUS_MSIZE(ptr)              ((msize_t *)(ptr))[-1]
 
 
 #define SMALL_PREVIOUS_MSIZE(ptr)              ((msize_t *)(ptr))[-1]
 

-#define SMALL_REGION_ADDRESS(region)   (((unsigned)region) << SMALL_BLOCKS_ALIGN)
-#define SMALL_REGION_END(region)       (SMALL_REGION_ADDRESS(region)+(1 << SMALL_BLOCKS_ALIGN))
+/*
+ * Convert between byte and msize units.
+ */
+#define SMALL_BYTES_FOR_MSIZE(_m)              ((_m) << SHIFT_SMALL_QUANTUM)
+#define SMALL_MSIZE_FOR_BYTES(_b)              ((_b) >> SHIFT_SMALL_QUANTUM)
+
+
+#define SMALL_REGION_ADDRESS(region)           ((unsigned char *)region)
+#define SMALL_REGION_END(region)               (SMALL_REGION_ADDRESS(region) + (NUM_SMALL_BLOCKS * SMALL_QUANTUM))

 
 

-typedef unsigned short small_region_t;
+/*
+ * Locate the heap base for a pointer known to be within a small region.
+ */
+#define SMALL_REGION_FOR_PTR(_p)               ((void *)((uintptr_t)(_p) & ~((1 << SMALL_BLOCKS_ALIGN) - 1)))
+
+/*
+ * Locate the metadata base for a pointer known to be within a small region.
+ */
+#define SMALL_HEADER_START                     (NUM_SMALL_BLOCKS * SMALL_QUANTUM)
+#define SMALL_META_HEADER_FOR_PTR(_p)          ((msize_t *)((uintptr_t)SMALL_REGION_FOR_PTR(_p) + SMALL_HEADER_START))

 
 

-#define INITIAL_NUM_SMALL_REGIONS      6       // must be even for szone to be aligned
+/*
+ * Compute the metadata index for a pointer known to be within a small region.
+ */
+#define SMALL_META_INDEX_FOR_PTR(_p)          (((uintptr_t)(_p) >> SHIFT_SMALL_QUANTUM) & (NUM_SMALL_CEIL_BLOCKS - 1))

 
 

-#define PROTECT_SMALL                  0       // Should be 0: 1 is too slow for normal use
+/*
+ * Find the metadata word for a pointer known to be within a small region.
+ */
+#define SMALL_METADATA_FOR_PTR(_p)             (SMALL_META_HEADER_FOR_PTR(_p) + SMALL_META_INDEX_FOR_PTR(_p))
+
+/*
+ * Determine whether a pointer known to be within a small region points to memory which is free.
+ */
+#define SMALL_PTR_IS_FREE(_p)                  (*SMALL_METADATA_FOR_PTR(_p) & SMALL_IS_FREE)
+
+/*
+ * Extract the msize value for a pointer known to be within a small region.
+ */
+#define SMALL_PTR_SIZE(_p)                     (*SMALL_METADATA_FOR_PTR(_p) & ~SMALL_IS_FREE)
+
+#define PROTECT_SMALL                          0       // Should be 0: 1 is too slow for normal use

 
 #define SMALL_CACHE    1
 #if !SMALL_CACHE
 #warning SMALL_CACHE turned off
 #endif
 
 
 #define SMALL_CACHE    1
 #if !SMALL_CACHE
 #warning SMALL_CACHE turned off
 #endif
 

-/********************* DEFINITIONS for large   ************************/
+/*********************  DEFINITIONS for large and huge  ***********************/

 
 #define LARGE_THRESHOLD                        (15 * 1024) // at or above this use "large"
 
 
 #define LARGE_THRESHOLD                        (15 * 1024) // at or above this use "large"
 


@@ -166,21 +329,38 @@ typedef unsigned short small_region_t;
     // But if the memory is only read, vm_copy() wins over memmove() at 3 or 4 pages (on a G3/300MHz)
     // This must be larger than LARGE_THRESHOLD
 
     // But if the memory is only read, vm_copy() wins over memmove() at 3 or 4 pages (on a G3/300MHz)
     // This must be larger than LARGE_THRESHOLD
 

+/*
+ * Given a large_entry, return the address of the allocated block.
+ */

 #define LARGE_ENTRY_ADDRESS(entry)                                     \
 #define LARGE_ENTRY_ADDRESS(entry)                                     \

-    (((entry).address_and_num_pages >> vm_page_shift) << vm_page_shift)
+    (void *)(((entry).address_and_num_pages >> vm_page_shift) << vm_page_shift)
+
+/*
+ * Given a large entry, return the number of pages or bytes in the allocated block.
+ */

 #define LARGE_ENTRY_NUM_PAGES(entry)                                   \
 #define LARGE_ENTRY_NUM_PAGES(entry)                                   \

-    ((entry).address_and_num_pages & ((1 << vm_page_shift) - 1))
+    ((entry).address_and_num_pages & (vm_page_size - 1))

 #define LARGE_ENTRY_SIZE(entry)                                                \
     (LARGE_ENTRY_NUM_PAGES(entry) << vm_page_shift)
 #define LARGE_ENTRY_SIZE(entry)                                                \
     (LARGE_ENTRY_NUM_PAGES(entry) << vm_page_shift)

+
+/*
+ * Compare a pointer with a large entry.
+ */

 #define LARGE_ENTRY_MATCHES(entry,ptr)                                 \
 #define LARGE_ENTRY_MATCHES(entry,ptr)                                 \

-    (!(((entry).address_and_num_pages - (unsigned)(ptr)) >> vm_page_shift))
-#define LARGE_ENTRY_IS_EMPTY(entry)    (!((entry).address_and_num_pages))
+    ((((entry).address_and_num_pages - (uintptr_t)(ptr)) >> vm_page_shift) == 0)
+
+#define LARGE_ENTRY_IS_EMPTY(entry)    (((entry).address_and_num_pages) == 0)

 
 typedef compact_range_t large_entry_t;
 
 typedef compact_range_t large_entry_t;

+typedef vm_range_t huge_entry_t;

 
 

-/********************* DEFINITIONS for huge    ************************/
+/*******************************************************************************
+ * Definitions for region hash
+ ******************************************************************************/

 
 

-typedef vm_range_t huge_entry_t;
+typedef void * region_t;
+
+#define INITIAL_NUM_REGIONS 63  // Must be odd to hash well

 
 /********************* zone itself     ************************/
 
 
 /********************* zone itself     ************************/
 


@@ -191,95 +371,242 @@ typedef struct {
     void               *log_address;
 
     /* Regions for tiny objects */
     void               *log_address;
 
     /* Regions for tiny objects */

-    unsigned           num_tiny_regions;
-    tiny_region_t      *tiny_regions;
+    size_t num_tiny_regions;
+    size_t num_tiny_regions_allocated;
+    region_t *tiny_regions;  // hashed by location
+    region_t *last_tiny_region;

     void               *last_tiny_free; // low SHIFT_TINY_QUANTUM indicate the msize
     unsigned           tiny_bitmap; // cache of the 32 free lists
     free_list_t                *tiny_free_list[NUM_TINY_SLOTS]; // 31 free lists for 1*TINY_QUANTUM to 31*TINY_QUANTUM plus 1 for larger than 32*SMALL_QUANTUM
     size_t             tiny_bytes_free_at_end; // the last free region in the last block is treated as a big block in use that is not accounted for
     unsigned           num_tiny_objects;
     void               *last_tiny_free; // low SHIFT_TINY_QUANTUM indicate the msize
     unsigned           tiny_bitmap; // cache of the 32 free lists
     free_list_t                *tiny_free_list[NUM_TINY_SLOTS]; // 31 free lists for 1*TINY_QUANTUM to 31*TINY_QUANTUM plus 1 for larger than 32*SMALL_QUANTUM
     size_t             tiny_bytes_free_at_end; // the last free region in the last block is treated as a big block in use that is not accounted for
     unsigned           num_tiny_objects;

-    unsigned           num_bytes_in_tiny_objects;
+    size_t             num_bytes_in_tiny_objects;

 
     /* Regions for small objects */
 
     /* Regions for small objects */

-    unsigned           num_small_regions;
-    small_region_t     *small_regions;
+    size_t num_small_regions;
+    size_t num_small_regions_allocated;
+    region_t *small_regions;  // hashed by location
+    region_t *last_small_region;

     void               *last_small_free; // low SHIFT_SMALL_QUANTUM indicate the msize
     unsigned           small_bitmap; // cache of the free list
     free_list_t                *small_free_list[NUM_SMALL_SLOTS];
     size_t             small_bytes_free_at_end; // the last free region in the last block is treated as a big block in use that is not accounted for
     unsigned           num_small_objects;
     void               *last_small_free; // low SHIFT_SMALL_QUANTUM indicate the msize
     unsigned           small_bitmap; // cache of the free list
     free_list_t                *small_free_list[NUM_SMALL_SLOTS];
     size_t             small_bytes_free_at_end; // the last free region in the last block is treated as a big block in use that is not accounted for
     unsigned           num_small_objects;

-    unsigned           num_bytes_in_small_objects;
+    size_t             num_bytes_in_small_objects;

 
     /* large objects: vm_page_shift <= log2(size) < 2 *vm_page_shift */
     unsigned           num_large_objects_in_use;
     unsigned           num_large_entries;
     large_entry_t      *large_entries; // hashed by location; null entries don't count
 
     /* large objects: vm_page_shift <= log2(size) < 2 *vm_page_shift */
     unsigned           num_large_objects_in_use;
     unsigned           num_large_entries;
     large_entry_t      *large_entries; // hashed by location; null entries don't count

-    unsigned           num_bytes_in_large_objects;
+    size_t             num_bytes_in_large_objects;

     
     /* huge objects: log2(size) >= 2 *vm_page_shift */
     
     /* huge objects: log2(size) >= 2 *vm_page_shift */

-    unsigned char      num_huge_entries;
+    unsigned           num_huge_entries;

     huge_entry_t       *huge_entries;
     huge_entry_t       *huge_entries;

-    unsigned           num_bytes_in_huge_objects;
+    size_t             num_bytes_in_huge_objects;

 
     /* Initial region list */
 
     /* Initial region list */

-    tiny_region_t      initial_tiny_regions[INITIAL_NUM_TINY_REGIONS];
-    small_region_t     initial_small_regions[INITIAL_NUM_SMALL_REGIONS];
+    region_t   initial_tiny_regions[INITIAL_NUM_REGIONS];
+    region_t   initial_small_regions[INITIAL_NUM_REGIONS];

 } szone_t;
 
 } szone_t;
 

-static void *szone_malloc(szone_t *szone, size_t size);
-static INLINE void *szone_malloc_should_clear(szone_t *szone, size_t size, boolean_t cleared_requested);
-static void szone_free(szone_t *szone, void *ptr);
-static boolean_t szone_check_all(szone_t *szone, const char *function);
-static void szone_print(szone_t *szone, boolean_t verbose);
-static void *small_malloc_from_region_no_lock(szone_t *szone, msize_t msize);
+#define SZONE_PAGED_SIZE       ((sizeof(szone_t) + vm_page_size - 1) & ~ (vm_page_size - 1))
+
+#if DEBUG_MALLOC || DEBUG_CLIENT
+static void            szone_sleep(void);
+#endif
+__private_extern__ void malloc_error_break(void);
+
+// msg prints after fmt, ...
+static void            szone_error(szone_t *szone, const char *msg, const void *ptr, const char *fmt, ...) __printflike(4, 5);
+
+static void            protect(void *address, size_t size, unsigned protection, unsigned debug_flags);
+static void            *allocate_pages(szone_t *szone, size_t size, unsigned char align, unsigned debug_flags, int vm_page_label);
+static void            deallocate_pages(szone_t *szone, void *addr, size_t size, unsigned debug_flags);
+static kern_return_t   _szone_default_reader(task_t task, vm_address_t address, vm_size_t size, void **ptr);
+
+static INLINE void     free_list_checksum(szone_t *szone, free_list_t *ptr, const char *msg) ALWAYSINLINE;
+static INLINE void     free_list_set_checksum(szone_t *szone, free_list_t *ptr) ALWAYSINLINE;
+static INLINE uintptr_t free_list_checksum_ptr(void *p) ALWAYSINLINE;
+static INLINE void * free_list_unchecksum_ptr(ptr_union ptr) ALWAYSINLINE;
+static unsigned                free_list_count(const free_list_t *ptr);
+
+static INLINE msize_t  get_tiny_meta_header(const void *ptr, boolean_t *is_free) ALWAYSINLINE;
+static INLINE void     set_tiny_meta_header_in_use(const void *ptr, msize_t msize) ALWAYSINLINE;
+static INLINE void     set_tiny_meta_header_middle(const void *ptr) ALWAYSINLINE;
+static INLINE void     set_tiny_meta_header_free(const void *ptr, msize_t msize) ALWAYSINLINE;
+static INLINE boolean_t        tiny_meta_header_is_free(const void *ptr) ALWAYSINLINE;
+static INLINE void     *tiny_previous_preceding_free(void *ptr, msize_t *prev_msize) ALWAYSINLINE;
+static void    tiny_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize);
+static void    tiny_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize);
+static INLINE region_t *tiny_region_for_ptr_no_lock(szone_t *szone, const void *ptr) ALWAYSINLINE;
+static INLINE void     tiny_free_no_lock(szone_t *szone, region_t *region, void *ptr, msize_t msize) ALWAYSINLINE;
+static void            *tiny_malloc_from_region_no_lock(szone_t *szone, msize_t msize);
+static INLINE boolean_t        try_realloc_tiny_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) ALWAYSINLINE;
+static boolean_t       tiny_check_region(szone_t *szone, region_t region);
+static kern_return_t   tiny_in_use_enumerator(task_t task, void *context, unsigned type_mask, szone_t *szone, memory_reader_t reader, vm_range_recorder_t recorder);
+static void    *tiny_malloc_from_free_list(szone_t *szone, msize_t msize);
+static INLINE void     *tiny_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested) ALWAYSINLINE;
+static INLINE void     free_tiny(szone_t *szone, void *ptr, region_t *tiny_region) ALWAYSINLINE;
+static void            print_tiny_free_list(szone_t *szone);
+static void            print_tiny_region(boolean_t verbose, region_t region, size_t bytes_at_end);
+static boolean_t       tiny_free_list_check(szone_t *szone, grain_t slot);
+
+static INLINE void     small_meta_header_set_is_free(msize_t *meta_headers, unsigned index, msize_t msize) ALWAYSINLINE;
+static INLINE void     small_meta_header_set_in_use(msize_t *meta_headers, msize_t index, msize_t msize) ALWAYSINLINE;
+static INLINE void     small_meta_header_set_middle(msize_t *meta_headers, msize_t index) ALWAYSINLINE;
+static void            small_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize);
+static void            small_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize);
+static INLINE region_t *small_region_for_ptr_no_lock(szone_t *szone, const void *ptr) ALWAYSINLINE;
+static INLINE void     small_free_no_lock(szone_t *szone, region_t *region, void *ptr, msize_t msize) ALWAYSINLINE;
+static void            *small_malloc_from_region_no_lock(szone_t *szone, msize_t msize);
+static INLINE boolean_t        try_realloc_small_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) ALWAYSINLINE;
+static boolean_t       szone_check_small_region(szone_t *szone, region_t region);
+static kern_return_t   small_in_use_enumerator(task_t task, void *context, unsigned type_mask, szone_t *szone, memory_reader_t reader, vm_range_recorder_t recorder);
+static void    *small_malloc_from_free_list(szone_t *szone, msize_t msize);
+static INLINE void     *small_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested) ALWAYSINLINE;
+static INLINE void     *small_malloc_cleared_no_lock(szone_t *szone, msize_t msize) ALWAYSINLINE;
+static INLINE void     free_small(szone_t *szone, void *ptr, region_t *small_region) ALWAYSINLINE;
+static void            print_small_free_list(szone_t *szone);
+static void            print_small_region(szone_t *szone, boolean_t verbose, region_t region, size_t bytes_at_end);
+static boolean_t       small_free_list_check(szone_t *szone, grain_t grain);
+
+static region_t * hash_lookup_region_no_lock(region_t *regions, size_t num_entries, region_t r);
+static void hash_region_insert_no_lock(region_t *regions, size_t num_entries, region_t r);
+static region_t * hash_regions_alloc_no_lock(szone_t *szone, size_t num_entries);
+static region_t * hash_regions_grow_no_lock(szone_t *szone, region_t *regions, size_t old_size, size_t *new_size);

 
 #if DEBUG_MALLOC
 
 #if DEBUG_MALLOC

-#define LOG(szone,ptr)                                                 \
-    (szone->log_address && (((unsigned)szone->log_address == -1) || (szone->log_address == (void *)(ptr))))
+static void            large_debug_print(szone_t *szone);
+#endif
+static large_entry_t   *large_entry_for_pointer_no_lock(szone_t *szone, const void *ptr);
+static void            large_entry_insert_no_lock(szone_t *szone, large_entry_t range);
+static INLINE void     large_entries_rehash_after_entry_no_lock(szone_t *szone, large_entry_t *entry) ALWAYSINLINE;
+static INLINE large_entry_t *large_entries_alloc_no_lock(szone_t *szone, unsigned num) ALWAYSINLINE;
+static void            large_entries_free_no_lock(szone_t *szone, large_entry_t *entries, unsigned num, vm_range_t *range_to_deallocate);
+static large_entry_t * large_entries_grow_no_lock(szone_t *szone, vm_range_t *range_to_deallocate);
+static vm_range_t      large_free_no_lock(szone_t *szone, large_entry_t *entry);
+static kern_return_t   large_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t large_entries_address, unsigned num_entries, memory_reader_t reader, vm_range_recorder_t recorder);
+static huge_entry_t    *huge_entry_for_pointer_no_lock(szone_t *szone, const void *ptr);
+static boolean_t       huge_entry_append(szone_t *szone, huge_entry_t huge);
+static kern_return_t   huge_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t huge_entries_address, unsigned num_entries, memory_reader_t reader, vm_range_recorder_t recorder);
+static void            *large_and_huge_malloc(szone_t *szone, size_t num_pages);
+static INLINE void     free_large_or_huge(szone_t *szone, void *ptr) ALWAYSINLINE;
+static INLINE int      try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) ALWAYSINLINE;
+
+static void            szone_free(szone_t *szone, void *ptr);
+static INLINE void     *szone_malloc_should_clear(szone_t *szone, size_t size, boolean_t cleared_requested) ALWAYSINLINE;
+static void            *szone_malloc(szone_t *szone, size_t size);
+static void            *szone_calloc(szone_t *szone, size_t num_items, size_t size);
+static void            *szone_valloc(szone_t *szone, size_t size);
+static size_t          szone_size(szone_t *szone, const void *ptr);
+static void            *szone_realloc(szone_t *szone, void *ptr, size_t new_size);
+static unsigned                szone_batch_malloc(szone_t *szone, size_t size, void **results, unsigned count);
+static void            szone_batch_free(szone_t *szone, void **to_be_freed, unsigned count);
+static void            szone_destroy(szone_t *szone);
+static size_t          szone_good_size(szone_t *szone, size_t size);
+
+static boolean_t       szone_check_all(szone_t *szone, const char *function);
+static boolean_t       szone_check(szone_t *szone);
+static kern_return_t   szone_ptr_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t zone_address, memory_reader_t reader, vm_range_recorder_t recorder);
+static void            szone_print(szone_t *szone, boolean_t verbose);
+static void            szone_log(malloc_zone_t *zone, void *log_address);
+static void            szone_force_lock(szone_t *szone);
+static void            szone_force_unlock(szone_t *szone);
+
+static void            szone_statistics(szone_t *szone, malloc_statistics_t *stats);
+
+static void            *frozen_malloc(szone_t *zone, size_t new_size);
+static void            *frozen_calloc(szone_t *zone, size_t num_items, size_t size);
+static void            *frozen_valloc(szone_t *zone, size_t new_size);
+static void            *frozen_realloc(szone_t *zone, void *ptr, size_t new_size);
+static void            frozen_free(szone_t *zone, void *ptr);
+static void            frozen_destroy(szone_t *zone);
+
+#if DEBUG_MALLOC
+# define LOG(szone,ptr)                                                        \
+    (szone->log_address && (((uintptr_t)szone->log_address == -1) || (szone->log_address == (void *)(ptr))))

 #else
 #else

-#define LOG(szone,ptr)         0
+# define LOG(szone,ptr)                0

 #endif
 
 #endif
 

-#define SZONE_LOCK(szone)              {       \
-    LOCK(szone->lock);                         \
-}
+#define SZONE_LOCK(szone)                                              \
+       do {                                                            \
+           LOCK(szone->lock);                                          \
+       } while (0)

 
 

-#define SZONE_UNLOCK(szone)            {       \
-    UNLOCK(szone->lock);                       \
-}
+#define SZONE_UNLOCK(szone)                                            \
+       do {                                                            \
+           UNLOCK(szone->lock);                                        \
+       } while (0)

 
 

-#define LOCK_AND_NOTE_LOCKED(szone,locked)     {       \
-    CHECK(szone, __PRETTY_FUNCTION__);                         \
-    locked = 1; SZONE_LOCK(szone);                     \
-}
+#define LOCK_AND_NOTE_LOCKED(szone,locked)                             \
+do {                                                                   \
+    CHECK(szone, __PRETTY_FUNCTION__);                                 \
+    locked = 1; SZONE_LOCK(szone);                                     \
+} while (0)

 
 #if DEBUG_MALLOC || DEBUG_CLIENT
 
 #if DEBUG_MALLOC || DEBUG_CLIENT

-#define CHECK(szone,fun)                                               \
+# define CHECK(szone,fun)                                              \

     if ((szone)->debug_flags & CHECK_REGIONS) szone_check_all(szone, fun)
 #else
     if ((szone)->debug_flags & CHECK_REGIONS) szone_check_all(szone, fun)
 #else

-#define CHECK(szone,fun)       {}
+# define CHECK(szone,fun)      do {} while (0)

 #endif
 
 /********************* VERY LOW LEVEL UTILITIES  ************************/
 
 #if DEBUG_MALLOC || DEBUG_CLIENT
 static void
 #endif
 
 /********************* VERY LOW LEVEL UTILITIES  ************************/
 
 #if DEBUG_MALLOC || DEBUG_CLIENT
 static void

-szone_sleep(void) {
+szone_sleep(void)
+{
+

     if (getenv("MallocErrorSleep")) {
     if (getenv("MallocErrorSleep")) {

-       malloc_printf("*** Sleeping to help debug\n");
+       _malloc_printf(ASL_LEVEL_NOTICE, "*** sleeping to help debug\n");

        sleep(3600); // to help debug
     }
 }
 #endif
 
        sleep(3600); // to help debug
     }
 }
 #endif
 

-static void
-szone_error(szone_t *szone, const char *msg, const void *ptr) {
+// msg prints after fmt, ...
+static __attribute__((noinline)) void
+szone_error(szone_t *szone, const char *msg, const void *ptr, const char *fmt, ...)
+{
+    va_list ap;
+    _SIMPLE_STRING b = _simple_salloc();
+

     if (szone) SZONE_UNLOCK(szone);
     if (szone) SZONE_UNLOCK(szone);

-    if (ptr) {
-       malloc_printf("*** malloc[%d]: error for object %p: %s\n", getpid(), ptr, msg);
+    if (b) {
+       if (fmt) {
+           va_start(ap, fmt);
+           _simple_vsprintf(b, fmt, ap);
+           va_end(ap);
+       }
+       if (ptr) {
+           _simple_sprintf(b, "*** error for object %p: %s\n", ptr, msg);
+       } else {
+           _simple_sprintf(b, "*** error: %s\n", msg);
+       }
+       malloc_printf("%s*** set a breakpoint in malloc_error_break to debug\n", _simple_string(b));
+       _simple_sfree(b);

     } else {
     } else {

-       malloc_printf("*** malloc[%d]: error: %s\n", getpid(), msg);
+       /*
+        * Should only get here if vm_allocate() can't get a single page of
+        * memory, implying _simple_asl_log() would also fail.  So we just
+        * print to the file descriptor.
+        */
+       if (fmt) {
+           va_start(ap, fmt);
+           _malloc_vprintf(MALLOC_PRINTF_NOLOG, fmt, ap);
+           va_end(ap);
+       }
+       if (ptr) {
+           _malloc_printf(MALLOC_PRINTF_NOLOG, "*** error for object %p: %s\n", ptr, msg);
+       } else {
+           _malloc_printf(MALLOC_PRINTF_NOLOG, "*** error: %s\n", msg);
+       }
+       _malloc_printf(MALLOC_PRINTF_NOLOG, "*** set a breakpoint in malloc_error_break to debug\n");

     }
     }

+    malloc_error_break();

 #if DEBUG_MALLOC
     szone_print(szone, 1);
     szone_sleep();
 #if DEBUG_MALLOC
     szone_print(szone, 1);
     szone_sleep();


@@ -287,126 +614,170 @@ szone_error(szone_t *szone, const char *msg, const void *ptr) {
 #if DEBUG_CLIENT
     szone_sleep();
 #endif
 #if DEBUG_CLIENT
     szone_sleep();
 #endif

+    if (szone->debug_flags & SCALABLE_MALLOC_ABORT_ON_ERROR) abort();

 }
 
 static void
 }
 
 static void

-protect(szone_t *szone, vm_address_t address, vm_size_t size,
-  unsigned protection, unsigned debug_flags) {
+protect(void *address, size_t size, unsigned protection, unsigned debug_flags)
+{

     kern_return_t      err;
     kern_return_t      err;

+

     if (!(debug_flags & SCALABLE_MALLOC_DONT_PROTECT_PRELUDE)) {
     if (!(debug_flags & SCALABLE_MALLOC_DONT_PROTECT_PRELUDE)) {

-       err = vm_protect(mach_task_self(), address - (1 << vm_page_shift), 1 << vm_page_shift,
-         0, protection);
+       err = vm_protect(mach_task_self(), (vm_address_t)(uintptr_t)address - vm_page_size, vm_page_size, 0, protection);

        if (err) {
        if (err) {

-           malloc_printf("*** malloc[%d]: Can't protect(%p) region for "
-             "prelude guard page at %p\n", getpid(), protection,
-             address - (1 << vm_page_shift));
+           malloc_printf("*** can't protect(%p) region for prelude guard page at %p\n",
+             protection,address - (1 << vm_page_shift));

        }
     }
     if (!(debug_flags & SCALABLE_MALLOC_DONT_PROTECT_POSTLUDE)) {
        }
     }
     if (!(debug_flags & SCALABLE_MALLOC_DONT_PROTECT_POSTLUDE)) {

-       err = vm_protect(mach_task_self(), (vm_address_t)(address + size), 1 << vm_page_shift, 0, protection);
+       err = vm_protect(mach_task_self(), (vm_address_t)(uintptr_t)address + size, vm_page_size, 0, protection);

        if (err) {
        if (err) {

-           malloc_printf("*** malloc[%d]: Can't protect(%p) region for "
-             "postlude guard page at %p\n", getpid(), protection,
-             address + size);
+           malloc_printf("*** can't protect(%p) region for postlude guard page at %p\n",
+             protection, address + size);

        }
     }
 }
 
        }
     }
 }
 

-static vm_address_t
-allocate_pages(szone_t *szone, size_t size, unsigned char align, unsigned debug_flags, int vm_page_label) {
+static void *
+allocate_pages(szone_t *szone, size_t size, unsigned char align, unsigned debug_flags, int vm_page_label)
+{

     // align specifies a desired alignment (as a log) or 0 if no alignment requested
     // align specifies a desired alignment (as a log) or 0 if no alignment requested

-    kern_return_t      err;
-    vm_address_t       addr;
+    void            *vm_addr;
+    uintptr_t          addr, aligned_address;

     boolean_t          add_guard_pages = debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES;
     size_t             allocation_size = round_page(size);
     boolean_t          add_guard_pages = debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES;
     size_t             allocation_size = round_page(size);

+    size_t             delta;
+    

     if (align) add_guard_pages = 0; // too cumbersome to deal with that
     if (!allocation_size) allocation_size = 1 << vm_page_shift;
     if (add_guard_pages) allocation_size += 2 * (1 << vm_page_shift);
     if (align) add_guard_pages = 0; // too cumbersome to deal with that
     if (!allocation_size) allocation_size = 1 << vm_page_shift;
     if (add_guard_pages) allocation_size += 2 * (1 << vm_page_shift);

-    if (align) allocation_size += 1 << align;
-    err = vm_allocate(mach_task_self(), &addr, allocation_size, vm_page_label | 1);
-    if (err) {
-       malloc_printf("*** malloc: vm_allocate(size=%d) failed (error code=%d)\n", size, err);
-       szone_error(szone, "Can't allocate region", NULL);
-       return NULL;
+    if (align) allocation_size += (size_t)1 << align;
+    vm_addr = mmap(0, allocation_size, PROT_READ | PROT_WRITE, MAP_ANON | MAP_PRIVATE, VM_MAKE_TAG(vm_page_label), 0);
+    if ((uintptr_t)vm_addr == -1) {
+       szone_error(szone, "can't allocate region", NULL, "*** mmap(size=%lld) failed (error code=%d)\n", (long long)allocation_size, errno);
+           return NULL;

     }
     }

+    addr = (uintptr_t)vm_addr;

     if (align) {
     if (align) {

-       // malloc_printf("In allocate_pages(size=%d(%p), align=%d) -> %p\n", size, size, align, addr);
-       vm_address_t    aligned_address = (addr + (1 << align) - 1) & ~ ((1 << align) - 1);
+       aligned_address = (addr + ((uintptr_t)1 << align) - 1) & ~ (((uintptr_t)1 << align) - 1);

        if (aligned_address != addr) {
        if (aligned_address != addr) {

-           size_t      delta = aligned_address - addr;
-           err = vm_deallocate(mach_task_self(), addr, delta);
-           if (err) malloc_printf("*** malloc: freeing unaligned header failed with %d\n", err);
-           // malloc_printf("deallocated unaligned header %p length=%d(%p)\n", addr, delta, delta);
+           delta = aligned_address - addr;
+           if (munmap((void *)addr, delta) == -1)
+               malloc_printf("*** freeing unaligned header failed with %d\n", errno);

            addr = aligned_address;
            allocation_size -= delta;
        }
        if (allocation_size > size) {
            addr = aligned_address;
            allocation_size -= delta;
        }
        if (allocation_size > size) {

-           err = vm_deallocate(mach_task_self(), addr+size, allocation_size - size);
-           if (err) malloc_printf("*** malloc: freeing unaligned footer failed with %d\n", err);
+           if (munmap((void *)(addr + size), allocation_size - size) == -1)
+               malloc_printf("*** freeing unaligned footer failed with %d\n", errno);

        }
     }
     if (add_guard_pages) {
        }
     }
     if (add_guard_pages) {

-       addr += 1 << vm_page_shift;
-       protect(szone, addr, size, 0, debug_flags);
+       addr += (uintptr_t)1 << vm_page_shift;
+       protect((void *)addr, size, 0, debug_flags);

     }
     }

-    return addr;
+    return (void *)addr;

 }
 
 static void
 }
 
 static void

-deallocate_pages(szone_t *szone, vm_address_t addr, size_t size, unsigned debug_flags) {
-    kern_return_t      err;
-    boolean_t          add_guard_pages = debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES;
+deallocate_pages(szone_t *szone, void *addr, size_t size, unsigned debug_flags)
+{
+    int        err;
+    boolean_t add_guard_pages = debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES;
+

     if (add_guard_pages) {
        addr -= 1 << vm_page_shift;
        size += 2 * (1 << vm_page_shift);
     }
     if (add_guard_pages) {
        addr -= 1 << vm_page_shift;
        size += 2 * (1 << vm_page_shift);
     }

-    err = vm_deallocate(mach_task_self(), addr, size);
-    if (err) {
-       szone_error(szone, "Can't deallocate_pages region", (void *)addr);
-    }
+    err = munmap(addr, size);
+    if ((err == -1) && szone)
+       szone_error(szone, "Can't deallocate_pages region", addr, NULL);

 }
 
 static kern_return_t
 }
 
 static kern_return_t

-_szone_default_reader(task_t task, vm_address_t address, vm_size_t size, void **ptr) {
+_szone_default_reader(task_t task, vm_address_t address, vm_size_t size, void **ptr)
+{

     *ptr = (void *)address;
     return 0;
 }
 
     *ptr = (void *)address;
     return 0;
 }
 

+/********************* FREE LIST UTILITIES  ************************/
+
+// A free list entry is comprised of a pair of pointers, previous and next.
+// Because the free list entries are previously freed objects, there is a
+// non-zero chance that a misbehaved program will write to an allocated object
+// after it has called free() on the pointer.  This write would then potentially
+// corrupt the previous and next pointers, leading to a crash.  In order to
+// detect this case, we take advantage of the fact that pointers are known to
+// be at least 16 byte aligned, and thus have at least 4 trailing zero bits.
+// When an entry is added to the free list, the previous and next pointers are
+// shifted right by 2 bits, and then have the high and low 2 bits set, to act
+// as guard bits.  Before accessing a free list object, we verify that these
+// bits are still set, and log an error if they are not.
+

 static INLINE void
 static INLINE void

-free_list_checksum(szone_t *szone, free_list_t *ptr, const char *msg) {
-    // We always checksum, as testing whether to do it (based on szone->debug_flags) is as fast as doing it
-    if (ptr->checksum != (((unsigned)ptr->previous) ^ ((unsigned)ptr->next) ^ CHECKSUM_MAGIC)) {
-#if DEBUG_MALLOC
-       malloc_printf("*** Incorrect checksum: %s\n", msg);
+free_list_checksum(szone_t *szone, free_list_t *ptr, const char *msg)
+{
+    uintptr_t ptrs = ptr->previous.u & ptr->next.u;
+
+#ifdef __LP64__
+    ptrs = (ptrs << 2) | (ptrs >> (64-2));
+#else
+    ptrs = (ptrs << 2) | (ptrs >> (32-2));
+#endif
+    
+    if ((ptrs & 15) != 15)
+       szone_error(szone, "incorrect checksum for freed object "
+                       "- object was probably modified after being freed.", ptr, NULL);
+}
+
+static INLINE uintptr_t
+free_list_checksum_ptr(void *p) 
+{
+    ptr_union ptr;
+    ptr.p = p;
+    
+#ifdef __LP64__
+    return (ptr.u >> 2) | 0xC000000000000003ULL;
+#else
+    return (ptr.u >> 2) | 0xC0000003U;

 #endif
 #endif

-       szone_error(szone, "Incorrect checksum for freed object - object was probably modified after being freed; break at szone_error", ptr);
-    }
+}
+
+static INLINE void *
+free_list_unchecksum_ptr(ptr_union ptr)
+{
+    uintptr_t u = (ptr.u >> 2) << 4;
+    return (void *)u;

 }
 
 static INLINE void
 }
 
 static INLINE void

-free_list_set_checksum(szone_t *szone, free_list_t *ptr) {
-    // We always set checksum, as testing whether to do it (based on
-    // szone->debug_flags) is slower than just doing it
-    ptr->checksum = ((unsigned)ptr->previous) ^ ((unsigned)ptr->next) ^ CHECKSUM_MAGIC;
+free_list_set_checksum(szone_t *szone, free_list_t *ptr)
+{
+    ptr->previous.u = free_list_checksum_ptr(ptr->previous.p);
+    ptr->next.u = free_list_checksum_ptr(ptr->next.p);

 }
 
 static unsigned
 }
 
 static unsigned

-free_list_count(const free_list_t *ptr) {
+free_list_count(const free_list_t *ptr)
+{

     unsigned   count = 0;
     unsigned   count = 0;

+

     while (ptr) {
     while (ptr) {

-       count++;
-//     malloc_printf("%p ", ptr);
-       ptr = ptr->next;
+        count++;
+        ptr = free_list_unchecksum_ptr(ptr->next);

     }
     return count;
 }
 
     }
     return count;
 }
 

+/* XXX inconsistent use of BITMAP32 and BITARRAY operations could be cleaned up */
+

 #define BITMAP32_SET(bitmap,bit)       (bitmap |= 1 << (bit))
 #define BITMAP32_CLR(bitmap,bit)       (bitmap &= ~ (1 << (bit)))
 #define BITMAP32_BIT(bitmap,bit)       ((bitmap >> (bit)) & 1)
 
 #define BITMAP32_SET(bitmap,bit)       (bitmap |= 1 << (bit))
 #define BITMAP32_CLR(bitmap,bit)       (bitmap &= ~ (1 << (bit)))
 #define BITMAP32_BIT(bitmap,bit)       ((bitmap >> (bit)) & 1)
 

-#define BITMAP32_FFS(bitmap) (ffs(bitmap))
-    // returns bit # of first bit that's one, starting at 1 (returns 0 if !bitmap)
+/* returns bit # of least-significant one bit, starting at 0 (undefined if !bitmap) */
+#define BITMAP32_CTZ(bitmap) (__builtin_ctz(bitmap))

 
 /********************* TINY FREE LIST UTILITIES        ************************/
 
 
 /********************* TINY FREE LIST UTILITIES        ************************/
 


@@ -424,352 +795,435 @@ free_list_count(const free_list_t *ptr) {
 #define BITARRAY_BIT(bits,index)       (((bits[index>>3]) >> (index & 7)) & 1)
 
 // Following is for start<8 and end<=start+32
 #define BITARRAY_BIT(bits,index)       (((bits[index>>3]) >> (index & 7)) & 1)
 
 // Following is for start<8 and end<=start+32

-#define BITARRAY_MCLR_LESS_32(bits,start,end) {        \
-    unsigned char      *_bits = (bits);        \
-    unsigned   _end = (end);                   \
-    switch (_end >> 3) {                       \
-       case 4: _bits[4] &= ~ ((1 << (_end - 32)) - 1); _end = 32;      \
-       case 3: _bits[3] &= ~ ((1 << (_end - 24)) - 1); _end = 24;      \
-       case 2: _bits[2] &= ~ ((1 << (_end - 16)) - 1); _end = 16;      \
-       case 1: _bits[1] &= ~ ((1 << (_end - 8)) - 1); _end = 8;        \
-       case 0: _bits[0] &= ~ ((1 << _end) - (1 << (start)));           \
-    }  \
-}
-
-#if 0  // Simple but slow version
-#warning Slow version in effect
-#define BITARRAY_MCLR(bits,index,num)  {               \
-    unsigned   _ctr = (num);                   \
-    unsigned   _cur = (index);                 \
-    while (_ctr--) {BITARRAY_CLR(bits,_cur); _cur++; } \
+static void ALWAYSINLINE
+bitarray_mclr(void *bits, unsigned start, unsigned end) {
+    unsigned word = OSReadLittleInt32(bits, 0);
+    unsigned mask = (0xFFFFFFFFU >> (31 - start)) >> 1;
+
+    if (end > 31) {
+        unsigned char *bytes = (unsigned char *)bits;
+        bytes[4] &= ~((1 << (end - 32)) - 1);
+    } else {
+        mask |= (0xFFFFFFFF << end);
+    }
+    OSWriteLittleInt32(bits, 0, word & mask);

 }
 }

-#else

 
 

-// Following is for num <= 32
-#define BITARRAY_MCLR(bits,index,num)  {               \
-    unsigned   _index = (index);                       \
-    unsigned char      *_rebased = (bits) + (_index >> 3);     \
-    _index &= 7;                                       \
-    BITARRAY_MCLR_LESS_32(_rebased, _index, _index + (num)); \
+/*
+ * Obtain the size of a free tiny block (in msize_t units).
+ */
+static msize_t
+get_tiny_free_size(const void *ptr)
+{
+    void *next_block = (void *)((uintptr_t)ptr + TINY_QUANTUM);
+    void *region_end = TINY_REGION_END(TINY_REGION_FOR_PTR(ptr));
+
+    // check whether the next block is outside the tiny region or a block header
+    // if so, then the size of this block is one, and there is no stored size.
+    if (next_block < region_end)
+    {
+        unsigned char *next_header = TINY_BLOCK_HEADER_FOR_PTR(next_block);
+        msize_t        next_index  = TINY_INDEX_FOR_PTR(next_block);
+        
+        if (!BITARRAY_BIT(next_header, next_index))
+            return TINY_FREE_SIZE(ptr);
+    }
+    return 1;
+}
+
+/*
+ * Get the size of the previous free block, which is stored in the last two
+ * bytes of the block.  If the previous block is not free, then the result is
+ * undefined.
+ */
+static msize_t
+get_tiny_previous_free_msize(const void *ptr)
+{
+    // check whether the previous block is in the tiny region and a block header
+    // if so, then the size of the previous block is one, and there is no stored
+    // size.
+    if (ptr != TINY_REGION_FOR_PTR(ptr))
+    {
+        void          *prev_block = (void *)((uintptr_t)ptr - TINY_QUANTUM);
+        unsigned char *prev_header = TINY_BLOCK_HEADER_FOR_PTR(prev_block);
+        msize_t        prev_index  = TINY_INDEX_FOR_PTR(prev_block);
+        if (BITARRAY_BIT(prev_header, prev_index))
+            return 1;
+        return TINY_PREVIOUS_MSIZE(ptr);
+    }
+    // don't read possibly unmapped memory before the beginning of the region
+    return 0;

 }
 }

-#endif

 
 static INLINE msize_t
 
 static INLINE msize_t

-get_tiny_meta_header(const void *ptr, boolean_t *is_free) {
+get_tiny_meta_header(const void *ptr, boolean_t *is_free)
+{

     // returns msize and is_free
     // may return 0 for the msize component (meaning 65536)
     // returns msize and is_free
     // may return 0 for the msize component (meaning 65536)

-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    msize_t            index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
-    unsigned           byte_index = index >> 3;
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+    unsigned           byte_index;
+
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    index = TINY_INDEX_FOR_PTR(ptr);
+    byte_index = index >> 3;
+      

     block_header += byte_index;
     index &= 7;
     *is_free = 0;
     block_header += byte_index;
     index &= 7;
     *is_free = 0;

-    if (!BITMAP32_BIT(*block_header, index)) return 0;
-    unsigned char      *in_use = block_header + (NUM_TINY_BLOCKS >> 3) + 4;
+    if (!BITMAP32_BIT(*block_header, index))
+        return 0;
+    in_use = TINY_INUSE_FOR_HEADER(block_header);

     if (!BITMAP32_BIT(*in_use, index)) {
     if (!BITMAP32_BIT(*in_use, index)) {

-       *is_free = 1;
-       return TINY_FREE_SIZE(ptr);
-    }
-#if defined(__BIG_ENDIAN__)
-    unsigned   *addr = (void *)((unsigned)block_header & ~3);
-    unsigned   word0 = OSReadSwapInt32(addr, 0);
-    unsigned   word1 = OSReadSwapInt32(addr, 4);
-    unsigned   bits = index + (((unsigned)block_header & 3) * 8);
-    unsigned   word = (word0 >> bits) | (word1 << (32 - bits));
-    unsigned   result = ffs(word >> 1);
-#if DEBUG_MALLOC
-    if (result >= 32) {
-       malloc_printf("*** get_tiny_meta_header() invariant broken %p %d\n", ptr, result);
-       szone_sleep();
-    }
-#endif
+        *is_free = 1;
+        return get_tiny_free_size(ptr);
+    }
+    uint32_t   *addr = (uint32_t *)((uintptr_t)block_header & ~3);
+    uint32_t   word0 = OSReadLittleInt32(addr, 0) >> index;
+    uint32_t   word1 = OSReadLittleInt32(addr, 4) << (8 - index);
+    uint32_t   bits = (((uintptr_t)block_header & 3) * 8);     // precision loss on LP64 OK here
+    uint32_t   word = (word0 >> bits) | (word1 << (24 - bits));
+    uint32_t   result = ffs(word >> 1);

     return result;
     return result;

-#else
-    unsigned   cur = index + 1;
-    while (!BITARRAY_BIT(block_header, cur)) cur++; // assumes padding at the zone end
-#if DEBUG_MALLOC
-    if (cur - index >= 32) {
-       malloc_printf("*** get_tiny_meta_header() invariant broken %p %d %d\n", ptr, index, cur);
-       szone_sleep();
-    }
-#endif
-    return cur - index;
-#endif

 }
 
 static INLINE void
 }
 
 static INLINE void

-set_tiny_meta_header_in_use(const void *ptr, msize_t msize) {
-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    msize_t            index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
+set_tiny_meta_header_in_use(const void *ptr, msize_t msize)
+{
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+    unsigned           byte_index;
+    msize_t            clr_msize;
+    unsigned           end_bit;
+       
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    index = TINY_INDEX_FOR_PTR(ptr);
+    byte_index = index >> 3;
+    

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-    if (msize >= 32) malloc_printf("*** set_tiny_meta_header_in_use() invariant broken %p %d\n", ptr, msize);
-    if ((unsigned)index + (unsigned)msize > 0x10000) malloc_printf("*** set_tiny_meta_header_in_use() invariant broken (2) %p %d\n", ptr, msize);
+    if (msize >= 32)
+       malloc_printf("set_tiny_meta_header_in_use() invariant broken %p %d\n", ptr, msize);
+    if ((unsigned)index + (unsigned)msize > 0x10000)
+       malloc_printf("set_tiny_meta_header_in_use() invariant broken (2) %p %d\n", ptr, msize);

 #endif
 #endif

-    unsigned           byte_index = index >> 3;

     block_header += byte_index;
     index &= 7;
     BITMAP32_SET(*block_header, index); 
     block_header += byte_index;
     index &= 7;
     BITMAP32_SET(*block_header, index); 

-    unsigned char      *in_use = block_header + (NUM_TINY_BLOCKS >> 3) + 4;
+    in_use = TINY_INUSE_FOR_HEADER(block_header);

     BITMAP32_SET(*in_use, index); 
     index++; 
     BITMAP32_SET(*in_use, index); 
     index++; 

-    msize_t    clr_msize = msize-1;
+    clr_msize = msize-1;

     if (clr_msize) {
        byte_index = index >> 3;
        block_header += byte_index; in_use += byte_index;
        index &= 7;
     if (clr_msize) {
        byte_index = index >> 3;
        block_header += byte_index; in_use += byte_index;
        index &= 7;

-       unsigned        end_bit = index + clr_msize;
-       BITARRAY_MCLR_LESS_32(block_header, index, end_bit);
-       BITARRAY_MCLR_LESS_32(in_use, index, end_bit);
+       end_bit = index + clr_msize;
+       bitarray_mclr(block_header, index, end_bit);
+       bitarray_mclr(in_use, index, end_bit);

     }
     BITARRAY_SET(block_header, index+clr_msize); // we set the block_header bit for the following block to reaffirm next block is a block
 #if DEBUG_MALLOC
     }
     BITARRAY_SET(block_header, index+clr_msize); // we set the block_header bit for the following block to reaffirm next block is a block
 #if DEBUG_MALLOC

-    boolean_t  ff;
-    msize_t    mf = get_tiny_meta_header(ptr, &ff);
-    if (msize != mf) {
-       malloc_printf("*** setting header for tiny in_use %p : %d\n", ptr, msize);
-       malloc_printf("reading header for tiny %p : %d %d\n", ptr, mf, ff);
+    {
+       boolean_t ff;
+       msize_t mf;
+       
+       mf = get_tiny_meta_header(ptr, &ff);
+       if (msize != mf) {
+           malloc_printf("setting header for tiny in_use %p : %d\n", ptr, msize);
+           malloc_printf("reading header for tiny %p : %d %d\n", ptr, mf, ff);
+       }

     }
 #endif
 }
 
 static INLINE void
     }
 #endif
 }
 
 static INLINE void

-set_tiny_meta_header_middle(const void *ptr) {
+set_tiny_meta_header_middle(const void *ptr)
+{

     // indicates this block is in the middle of an in use block
     // indicates this block is in the middle of an in use block

-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    unsigned char      *in_use = (unsigned char *)(headers_start + (NUM_TINY_BLOCKS >> 3) + 4);
-    msize_t            index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
-    BITARRAY_CLR(block_header, index); BITARRAY_CLR(in_use, index); 
-    TINY_FREE_SIZE(ptr) = 0;
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    in_use = TINY_INUSE_FOR_HEADER(block_header);
+    index = TINY_INDEX_FOR_PTR(ptr);
+
+    BITARRAY_CLR(block_header, index);
+    BITARRAY_CLR(in_use, index); 

 }
 
 static INLINE void
 }
 
 static INLINE void

-set_tiny_meta_header_free(const void *ptr, msize_t msize) {
+set_tiny_meta_header_free(const void *ptr, msize_t msize)
+{

     // !msize is acceptable and means 65536
     // !msize is acceptable and means 65536

-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    unsigned char      *in_use = (unsigned char *)(headers_start + (NUM_TINY_BLOCKS >> 3) + 4);
-    msize_t            index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    in_use = TINY_INUSE_FOR_HEADER(block_header);
+    index = TINY_INDEX_FOR_PTR(ptr);
+

 #if DEBUG_MALLOC
     if ((unsigned)index + (unsigned)msize > 0x10000) {
 #if DEBUG_MALLOC
     if ((unsigned)index + (unsigned)msize > 0x10000) {

-       malloc_printf("*** setting header for tiny free %p msize too large: %d\n", ptr, msize);
+       malloc_printf("setting header for tiny free %p msize too large: %d\n", ptr, msize);

     }
 #endif
     }
 #endif

-    BITARRAY_SET(block_header, index); BITARRAY_CLR(in_use, index); 
-    TINY_FREE_SIZE(ptr) = msize;
-    // mark the end of this block
-    if (msize) {       // msize==0 => the whole region is free
-       void    *follower = FOLLOWING_TINY_PTR(ptr, msize);
-       TINY_PREVIOUS_MSIZE(follower) = msize;
+    BITARRAY_SET(block_header, index);
+    BITARRAY_CLR(in_use, index); 
+    // mark the end of this block if msize is > 1.  For msize == 0, the whole
+    // region is free, so there is no following block. For msize == 1, there is
+    // no space to write the size on 64 bit systems.  The size for 1 quantum
+    // blocks is computed from the metadata bitmaps.
+    if (msize > 1) {
+        void   *follower = FOLLOWING_TINY_PTR(ptr, msize);
+        TINY_PREVIOUS_MSIZE(follower) = msize;
+        TINY_FREE_SIZE(ptr) = msize;
+    }
+    if (msize == 0) {
+        TINY_FREE_SIZE(ptr) = msize;

     }
 #if DEBUG_MALLOC
     boolean_t  ff;
     msize_t    mf = get_tiny_meta_header(ptr, &ff);
     if ((msize != mf) || !ff) {
     }
 #if DEBUG_MALLOC
     boolean_t  ff;
     msize_t    mf = get_tiny_meta_header(ptr, &ff);
     if ((msize != mf) || !ff) {

-       malloc_printf("*** setting header for tiny free %p : %d\n", ptr, msize);
-       malloc_printf("reading header for tiny %p : %d %d\n", ptr, mf, ff);
+       malloc_printf("setting header for tiny free %p : %u\n", ptr, msize);
+       malloc_printf("reading header for tiny %p : %u %u\n", ptr, mf, ff);

     }
 #endif
 }
 
 static INLINE boolean_t
     }
 #endif
 }
 
 static INLINE boolean_t

-tiny_meta_header_is_free(const void *ptr) {
-    // returns msize and is_free shifted by 16
-    // may return 0 for the msize component (meaning 65536)
-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    unsigned char      *in_use = (unsigned char *)(headers_start + (NUM_TINY_BLOCKS >> 3) + 4);
-    msize_t            index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
-    if (!BITARRAY_BIT(block_header, index)) return 0;
+tiny_meta_header_is_free(const void *ptr)
+{
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    in_use = TINY_INUSE_FOR_HEADER(block_header);
+    index = TINY_INDEX_FOR_PTR(ptr);
+    if (!BITARRAY_BIT(block_header, index))
+       return 0;

     return !BITARRAY_BIT(in_use, index);
 }
 
 static INLINE void *
     return !BITARRAY_BIT(in_use, index);
 }
 
 static INLINE void *

-tiny_previous_preceding_free(void *ptr, msize_t *prev_msize) {
+tiny_previous_preceding_free(void *ptr, msize_t *prev_msize)
+{

     // returns the previous block, assuming and verifying it's free
     // returns the previous block, assuming and verifying it's free

-    unsigned short     shifted_base = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << TINY_BLOCKS_ALIGN;
-    unsigned char      *block_header = (unsigned char *)headers_start;
-    unsigned char      *in_use = (unsigned char *)(headers_start + (NUM_TINY_BLOCKS >> 3) + 4);
-    msize_t    index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
-    if (!index) return NULL;
-    msize_t    previous_msize = TINY_PREVIOUS_MSIZE(ptr);
-    if (previous_msize > index) return NULL;
-    msize_t    previous_index = index - previous_msize;
-    void       *previous_ptr = (void *)((shifted_base << TINY_BLOCKS_ALIGN) + (previous_index << SHIFT_TINY_QUANTUM));
-    if (TINY_FREE_SIZE(previous_ptr) != previous_msize) return NULL;
-    if (!BITARRAY_BIT(block_header, previous_index)) return NULL;
-    if (BITARRAY_BIT(in_use, previous_index)) return NULL;
+    unsigned char      *block_header;
+    unsigned char      *in_use;
+    msize_t            index;
+    msize_t            previous_msize;
+    msize_t            previous_index;
+    void               *previous_ptr;
+
+    block_header = TINY_BLOCK_HEADER_FOR_PTR(ptr);
+    in_use = TINY_INUSE_FOR_HEADER(block_header);
+    index = TINY_INDEX_FOR_PTR(ptr);
+    
+    if (!index)
+        return NULL;
+    if ((previous_msize = get_tiny_previous_free_msize(ptr)) > index)
+        return NULL;
+    
+    previous_index = index - previous_msize;
+    previous_ptr = (void *)(TINY_REGION_FOR_PTR(ptr) + TINY_BYTES_FOR_MSIZE(previous_index));
+    if (!BITARRAY_BIT(block_header, previous_index))
+        return NULL;
+    if (BITARRAY_BIT(in_use, previous_index))
+        return NULL;
+    if (get_tiny_free_size(previous_ptr) != previous_msize)
+        return NULL;
+    

     // conservative check did match true check
     *prev_msize = previous_msize;
     // conservative check did match true check
     *prev_msize = previous_msize;

-    // malloc_printf("tiny_previous_preceding_free(%p) -> %p,%d\n", ptr, previous_ptr, previous_msize);

     return previous_ptr;
 }
 
     return previous_ptr;
 }
 

-static INLINE void
-tiny_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize) {
-    // Adds an item to the proper free list
-    // Also marks the meta-header of the block properly
-    // Assumes szone has been locked
+/* 
+ * Adds an item to the proper free list, and also marks the meta-header of the 
+ * block properly.
+ * Assumes szone has been locked
+ */ 
+static void
+tiny_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize)
+{

     grain_t    slot = (!msize || (msize >= NUM_TINY_SLOTS)) ? NUM_TINY_SLOTS - 1 : msize - 1;
     free_list_t        *free_ptr = ptr;
     free_list_t        *free_head = szone->tiny_free_list[slot];
     grain_t    slot = (!msize || (msize >= NUM_TINY_SLOTS)) ? NUM_TINY_SLOTS - 1 : msize - 1;
     free_list_t        *free_ptr = ptr;
     free_list_t        *free_head = szone->tiny_free_list[slot];

+

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In tiny_free_list_add_ptr(), ptr=%p, msize=%d\n", ptr, msize);
+        malloc_printf("in %s, ptr=%p, msize=%d\n", __FUNCTION__, ptr, msize);

     }
     }

-    if (((unsigned)ptr) & (TINY_QUANTUM - 1)) {
-       szone_error(szone, "tiny_free_list_add_ptr: Unaligned ptr", ptr);
+    if (((uintptr_t)ptr) & (TINY_QUANTUM - 1)) {
+        szone_error(szone, "tiny_free_list_add_ptr: Unaligned ptr", ptr, NULL);

     }
 #endif
     set_tiny_meta_header_free(ptr, msize);
     if (free_head) {
     }
 #endif
     set_tiny_meta_header_free(ptr, msize);
     if (free_head) {

-       free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);
+        free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (free_head->previous) {
-           malloc_printf("ptr=%p slot=%d free_head=%p previous=%p\n", ptr, slot, free_head, free_head->previous);
-           szone_error(szone, "tiny_free_list_add_ptr: Internal invariant broken (free_head->previous)", ptr);
-       }
-       if (! tiny_meta_header_is_free(free_head)) {
-           malloc_printf("ptr=%p slot=%d free_head=%p\n", ptr, slot, free_head);
-           szone_error(szone, "tiny_free_list_add_ptr: Internal invariant broken (free_head is not a free pointer)", ptr);
-       }
+        if (free_list_unchecksum_ptr(free_head->previous)) {
+            szone_error(szone, "tiny_free_list_add_ptr: Internal invariant broken (free_head->previous)", ptr,
+               "ptr=%p slot=%d free_head=%p previous=%p\n", ptr, slot, free_head, free_head->previous.p);
+        }
+        if (! tiny_meta_header_is_free(free_head)) {
+            szone_error(szone, "tiny_free_list_add_ptr: Internal invariant broken (free_head is not a free pointer)", ptr,
+               "ptr=%p slot=%d free_head=%p\n", ptr, slot, free_head);
+        }

 #endif
 #endif

-       free_head->previous = free_ptr;
-       free_list_set_checksum(szone, free_head);
+        free_head->previous.u = free_list_checksum_ptr(free_ptr);

     } else {
     } else {

-       BITMAP32_SET(szone->tiny_bitmap, slot);
+        BITMAP32_SET(szone->tiny_bitmap, slot);

     }
     }

-    free_ptr->previous = NULL;
-    free_ptr->next = free_head;
+    free_ptr->previous.p = NULL;
+    free_ptr->next.p = free_head;

     free_list_set_checksum(szone, free_ptr);
     szone->tiny_free_list[slot] = free_ptr;
     free_list_set_checksum(szone, free_ptr);
     szone->tiny_free_list[slot] = free_ptr;

-    // malloc_printf("Setting head of free list for slot=%d to %p\n", slot, free_ptr);

 }
 
 }
 

+/* 
+ * Removes the item pointed to by ptr in the proper free list. 
+ * Assumes szone has been locked
+ */ 

 static INLINE void
 static INLINE void

-tiny_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize) {
-    // Removes item in the proper free list
-    // msize could be read, but all callers have it so we pass it in
-    // Assumes szone has been locked
+tiny_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize)
+{

     grain_t    slot = (!msize || (msize >= NUM_TINY_SLOTS)) ? NUM_TINY_SLOTS - 1 : msize - 1;
     grain_t    slot = (!msize || (msize >= NUM_TINY_SLOTS)) ? NUM_TINY_SLOTS - 1 : msize - 1;

-    free_list_t        *free_ptr = ptr;
-    free_list_t        *next = free_ptr->next;
-    free_list_t        *previous = free_ptr->previous;
+    free_list_t        *free_ptr = ptr, *next, *previous;
+    free_list_checksum(szone, free_ptr, __PRETTY_FUNCTION__);
+
+    next = free_list_unchecksum_ptr(free_ptr->next);
+    previous = free_list_unchecksum_ptr(free_ptr->previous);
+

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In tiny_free_list_remove_ptr(), ptr=%p, msize=%d\n", ptr, msize);
+        malloc_printf("In %s, ptr=%p, msize=%d\n", __FUNCTION__, ptr, msize);

     }
 #endif
     }
 #endif

-    free_list_checksum(szone, free_ptr, __PRETTY_FUNCTION__);
-    if (!previous) {
-       // The block to remove is the head of the free list
+    if (!previous) { 
+    // The block to remove is the head of the free list

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (szone->tiny_free_list[slot] != ptr) {
-           malloc_printf("ptr=%p slot=%d msize=%d szone->tiny_free_list[slot]=%p\n", ptr, slot, msize, szone->tiny_free_list[slot]);
-           szone_error(szone, "tiny_free_list_remove_ptr: Internal invariant broken (szone->tiny_free_list[slot])", ptr);
-           return;
-       }
+        if (szone->tiny_free_list[slot] != ptr) {
+            szone_error(szone, "tiny_free_list_remove_ptr: Internal invariant broken (szone->tiny_free_list[slot])", ptr,
+              "ptr=%p slot=%d msize=%d szone->tiny_free_list[slot]=%p\n",
+              ptr, slot, msize, szone->tiny_free_list[slot]);
+            return;
+        }

 #endif
 #endif

-       szone->tiny_free_list[slot] = next;
-       if (!next) BITMAP32_CLR(szone->tiny_bitmap, slot);
+        szone->tiny_free_list[slot] = next;
+        if (!next) BITMAP32_CLR(szone->tiny_bitmap, slot);

     } else {
     } else {

-       previous->next = next;
-       free_list_set_checksum(szone, previous);
+        // We know free_ptr is already checksummed, so we don't need to do it
+        // again.
+        previous->next = free_ptr->next;

     }
     if (next) {
     }
     if (next) {

-       next->previous = previous;
-       free_list_set_checksum(szone, next);
+        // We know free_ptr is already checksummed, so we don't need to do it
+        // again.
+        next->previous = free_ptr->previous;

     }
 }
 
     }
 }
 

-static INLINE tiny_region_t *
-tiny_region_for_ptr_no_lock(szone_t *szone, const void *ptr) {
-    tiny_region_t      *region = szone->tiny_regions;
-    unsigned           num_regions = szone->num_tiny_regions;
-    unsigned           ptr_shifted = ((unsigned)ptr) >> TINY_BLOCKS_ALIGN;
-    while (num_regions--) {
-       tiny_region_t   this = *region;
-       if (ptr_shifted == this) return region;
-       region++;
-    }
-    return NULL;
+/*
+ * tiny_region_for_ptr_no_lock - Returns the tiny region containing the pointer,
+ * or NULL if not found.
+ */
+static INLINE region_t *
+tiny_region_for_ptr_no_lock(szone_t *szone, const void *ptr)
+{
+  return hash_lookup_region_no_lock(szone->tiny_regions,
+                                    szone->num_tiny_regions_allocated,
+                                    TINY_REGION_FOR_PTR(ptr));

 }
 
 static INLINE void
 }
 
 static INLINE void

-tiny_free_no_lock(szone_t *szone, tiny_region_t *region, void *ptr, msize_t msize) {
-    size_t     original_size = msize << SHIFT_TINY_QUANTUM;
+tiny_free_no_lock(szone_t *szone, region_t *region, void *ptr, msize_t msize)
+{
+    size_t     original_size = TINY_BYTES_FOR_MSIZE(msize);

     void       *next_block = ((char *)ptr + original_size);
     void       *next_block = ((char *)ptr + original_size);

+    msize_t    previous_msize;
+    void       *previous;
+    msize_t    next_msize;
+    free_list_t        *big_free_block;
+    free_list_t        *after_next_block;
+    free_list_t        *before_next_block;
+

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In tiny_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);
+       malloc_printf("in tiny_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);

     }
     if (! msize) {
     }
     if (! msize) {

-       malloc_printf("In tiny_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);
-       szone_error(szone, "Trying to free tiny block that is too small", ptr);
+       szone_error(szone, "trying to free tiny block that is too small", ptr,
+           "in tiny_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);

     }
 #endif
     // We try to coalesce this block with the preceeding one
     }
 #endif
     // We try to coalesce this block with the preceeding one

-    msize_t    previous_msize;
-    void       *previous = tiny_previous_preceding_free(ptr, &previous_msize);
+    previous = tiny_previous_preceding_free(ptr, &previous_msize);

     if (previous) {
 #if DEBUG_MALLOC
     if (previous) {
 #if DEBUG_MALLOC

-       if (LOG(szone, ptr) || LOG(szone,previous)) { 
-           malloc_printf("In tiny_free_no_lock(), coalesced backwards for %p previous=%p\n", ptr, previous);
-       }
+        if (LOG(szone, ptr) || LOG(szone,previous)) { 
+            malloc_printf("in tiny_free_no_lock(), coalesced backwards for %p previous=%p\n", ptr, previous);
+        }

 #endif
 #endif

-       tiny_free_list_remove_ptr(szone, previous, previous_msize);
-       ptr = previous;
-       msize += previous_msize;
+        // clear the meta_header since this is no longer the start of a block
+        set_tiny_meta_header_middle(ptr);
+        tiny_free_list_remove_ptr(szone, previous, previous_msize);
+        ptr = previous;
+        msize += previous_msize;

     }
     // We try to coalesce with the next block
     }
     // We try to coalesce with the next block

-    if (((vm_address_t)next_block < TINY_REGION_END(*region)) && tiny_meta_header_is_free(next_block)) {
-       // The next block is free, we coalesce
-       msize_t next_msize = TINY_FREE_SIZE(next_block);
+    if ((next_block < TINY_REGION_END(*region)) && tiny_meta_header_is_free(next_block)) {
+        next_msize = get_tiny_free_size(next_block);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone, ptr) || LOG(szone, next_block)) {
-           malloc_printf("In tiny_free_no_lock(), for ptr=%p, msize=%d coalesced forward=%p next_msize=%d\n", ptr, msize, next_block, next_msize);
-       }
+        if (LOG(szone, ptr) || LOG(szone, next_block)) {
+            malloc_printf("in tiny_free_no_lock(), for ptr=%p, msize=%d coalesced forward=%p next_msize=%d\n",
+              ptr, msize, next_block, next_msize);
+        }

 #endif
 #endif

-       if (next_msize >= NUM_TINY_SLOTS) {
-           // we take a short cut here to avoid removing next_block from the slot 31 freelist and then adding ptr back to slot 31
-           // malloc_printf("Replacing %p(msize=%d) with %p(msize=%d) in freelist\n", next_block, next_msize, ptr, msize+next_msize);
-           msize += next_msize;
-           free_list_t *big_free_block = (free_list_t *)next_block;
-           free_list_t *after_next_block = big_free_block->next;
-           free_list_t *before_next_block = big_free_block->previous;
-           free_list_checksum(szone, big_free_block, __PRETTY_FUNCTION__);
-           if (!before_next_block) {
-               szone->tiny_free_list[NUM_TINY_SLOTS-1] = ptr;
-           } else {
-               before_next_block->next = ptr;
-               free_list_set_checksum(szone, before_next_block);
-           }
-           if (after_next_block) {
-               after_next_block->previous = ptr;
-               free_list_set_checksum(szone, after_next_block);
-           }
-           ((free_list_t *)ptr)->previous = before_next_block;
-           ((free_list_t *)ptr)->next = after_next_block;
-           free_list_set_checksum(szone, ptr);
-           set_tiny_meta_header_free(ptr, msize);
-           set_tiny_meta_header_middle(big_free_block); // clear the meta_header to enable coalescing backwards
-           goto tiny_free_ending;
-       }
-       tiny_free_list_remove_ptr(szone, next_block, next_msize);
-       set_tiny_meta_header_middle(next_block); // clear the meta_header to enable coalescing backwards
-       msize += next_msize;
-    }
+        // If we are coalescing with the next block, and the next block is in
+        // the last slot of the free list, then we optimize this case here to 
+        // avoid removing next_block from the slot 31 and then adding ptr back
+        // to slot 31.
+        if (next_msize >= NUM_TINY_SLOTS) {
+            msize += next_msize;
+            big_free_block = (free_list_t *)next_block;
+            free_list_checksum(szone, big_free_block, __PRETTY_FUNCTION__);
+            after_next_block = free_list_unchecksum_ptr(big_free_block->next);
+            before_next_block = free_list_unchecksum_ptr(big_free_block->previous);
+            if (!before_next_block) {
+                szone->tiny_free_list[NUM_TINY_SLOTS-1] = ptr;
+            } else {
+                before_next_block->next.u = free_list_checksum_ptr(ptr);
+            }
+            if (after_next_block) {
+                after_next_block->previous.u = free_list_checksum_ptr(ptr);
+            }
+            // we don't need to checksum these since they are already checksummed
+            ((free_list_t *)ptr)->previous = big_free_block->previous;
+            ((free_list_t *)ptr)->next = big_free_block->next;
+            
+            // clear the meta_header to enable coalescing backwards
+            set_tiny_meta_header_middle(big_free_block);
+            set_tiny_meta_header_free(ptr, msize);
+            goto tiny_free_ending;
+        }
+        tiny_free_list_remove_ptr(szone, next_block, next_msize);
+        set_tiny_meta_header_middle(next_block); // clear the meta_header to enable coalescing backwards
+        msize += next_msize;
+    }
+#if !TINY_CACHE
+    // The tiny cache already scribbles free blocks as they go through the
+    // cache, so we do not need to do it here.

     if ((szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) && msize) {
     if ((szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) && msize) {

-       memset(ptr, 0x55, msize << SHIFT_TINY_QUANTUM);
+       memset(ptr, 0x55, TINY_BYTES_FOR_MSIZE(msize));

     }
     }

+#endif

     tiny_free_list_add_ptr(szone, ptr, msize);
   tiny_free_ending:
     // When in proper debug mode we write on the memory to help debug memory smashers
     tiny_free_list_add_ptr(szone, ptr, msize);
   tiny_free_ending:
     // When in proper debug mode we write on the memory to help debug memory smashers


@@ -777,161 +1231,245 @@ tiny_free_no_lock(szone_t *szone, tiny_region_t *region, void *ptr, msize_t msiz
     szone->num_bytes_in_tiny_objects -= original_size; // we use original_size and not msize to avoid double counting the coalesced blocks
 }
 
     szone->num_bytes_in_tiny_objects -= original_size; // we use original_size and not msize to avoid double counting the coalesced blocks
 }
 

+// Allocates from the last region or a freshly allocated region

 static void *
 static void *

-tiny_malloc_from_region_no_lock(szone_t *szone, msize_t msize) {
-    // Allocates from the last region or a freshly allocated region
-    // Before anything we transform the tiny_bytes_free_at_end - if any - to a regular free block
+tiny_malloc_from_region_no_lock(szone_t *szone, msize_t msize)
+{
+    void            *last_block, *ptr, *aligned_address;
+    unsigned char   *last_header;
+    msize_t         last_msize, last_index;
+
+    // Before anything we transform any remaining tiny_bytes_free_at_end into a
+    // regular free block.  We take special care here to update the bitfield
+    // information, since we are bypassing the normal free codepath.  If there
+    // is more than one quanta worth of memory in tiny_bytes_free_at_end, then
+    // there will be two block headers:
+    // 1) header for the free space at end, msize = 1
+    // 2) header inserted by set_tiny_meta_header_in_use after block
+    // We must clear the second one so that when the free block's size is
+    // queried, we do not think the block is only 1 quantum in size because
+    // of the second set header bit.

     if (szone->tiny_bytes_free_at_end) {
     if (szone->tiny_bytes_free_at_end) {

-       tiny_region_t   last_region = szone-> tiny_regions[szone->num_tiny_regions-1];
-       void    *last_block = (void *)(TINY_REGION_END(last_region) - szone->tiny_bytes_free_at_end);
-       tiny_free_list_add_ptr(szone, last_block, szone->tiny_bytes_free_at_end >> SHIFT_TINY_QUANTUM);
-       szone->tiny_bytes_free_at_end = 0;
+        last_block = TINY_REGION_END(szone->last_tiny_region) - szone->tiny_bytes_free_at_end;
+        last_msize = TINY_MSIZE_FOR_BYTES(szone->tiny_bytes_free_at_end);
+        last_header = TINY_BLOCK_HEADER_FOR_PTR(last_block);
+        last_index  = TINY_INDEX_FOR_PTR(last_block);
+        
+        if (last_index != (NUM_TINY_BLOCKS - 1))
+            BITARRAY_CLR(last_header, last_index + 1);
+        
+        tiny_free_list_add_ptr(szone, last_block, last_msize);
+        szone->tiny_bytes_free_at_end = 0;

     }
     }

-    void               *ptr;

     // time to create a new region
     // time to create a new region

-    vm_address_t       aligned_address = allocate_pages(szone, TINY_REGION_SIZE, TINY_BLOCKS_ALIGN, 0, VM_MAKE_TAG(VM_MEMORY_MALLOC_TINY));
-    if (! aligned_address) {
-       // out of memory!
-       return NULL;
-    }
-    // malloc_printf("Allocated tiny region #%d: %p [%y]\n", szone->num_tiny_regions, aligned_address, TINY_REGION_SIZE);
+    aligned_address = allocate_pages(szone, TINY_REGION_SIZE, TINY_BLOCKS_ALIGN, 0, VM_MEMORY_MALLOC_TINY);
+    if (!aligned_address) // out of memory!
+        return NULL;

     // We set the padding after block_header to be all 1
     // We set the padding after block_header to be all 1

-    ((unsigned *)(aligned_address + (1 << TINY_BLOCKS_ALIGN) + (NUM_TINY_BLOCKS >> 3)))[0] = ~0;
-    if (szone->num_tiny_regions == INITIAL_NUM_TINY_REGIONS) {
-       tiny_region_t   *new_regions;
-       // malloc_printf("=== Growing tiny_regions (%d regions)\n", szone->num_tiny_regions);
-       new_regions = small_malloc_from_region_no_lock(szone, 16); // 16 * 512 bytes is plenty of tiny regions (more than 4,000)
-       if (!new_regions) return NULL;
-       memcpy(new_regions, szone->tiny_regions, INITIAL_NUM_TINY_REGIONS * sizeof(tiny_region_t));
-       szone->tiny_regions = new_regions; // we set the pointer after it's all ready to enable enumeration from another thread without locking
-    }
-    szone->tiny_regions[szone->num_tiny_regions] = aligned_address >> TINY_BLOCKS_ALIGN;
-    szone->num_tiny_regions ++; // we set the number after the pointer is all ready to enable enumeration from another thread without taking the lock
-    ptr = (void *)aligned_address; 
+    ((uint32_t *)(aligned_address + TINY_HEADER_START + (NUM_TINY_BLOCKS >> 3)))[0] = ~0;
+
+    // Check to see if the hash ring of tiny regions needs to grow.  Try to
+    // avoid the hash ring becoming too dense.
+    if (szone->num_tiny_regions_allocated < (2 * szone->num_tiny_regions)) {
+      region_t *new_regions;
+      size_t new_size;
+      new_regions = hash_regions_grow_no_lock(szone, szone->tiny_regions,
+                                              szone->num_tiny_regions_allocated,
+                                              &new_size);
+      // Do not deallocate the current tiny_regions allocation since someone may
+         // be iterating it.  Instead, just leak it.
+      szone->tiny_regions = new_regions;
+      szone->num_tiny_regions_allocated = new_size;
+    }
+    // Insert the new region into the hash ring, and update malloc statistics
+    hash_region_insert_no_lock(szone->tiny_regions, 
+                               szone->num_tiny_regions_allocated, 
+                               aligned_address);
+    szone->last_tiny_region = aligned_address;
+
+    szone->num_tiny_regions++;
+    ptr = aligned_address; 

     set_tiny_meta_header_in_use(ptr, msize);
     szone->num_tiny_objects++;
     set_tiny_meta_header_in_use(ptr, msize);
     szone->num_tiny_objects++;

-    szone->num_bytes_in_tiny_objects += msize << SHIFT_TINY_QUANTUM;
+    szone->num_bytes_in_tiny_objects += TINY_BYTES_FOR_MSIZE(msize);
+ 

     // We put a header on the last block so that it appears in use (for coalescing, etc...)
     // We put a header on the last block so that it appears in use (for coalescing, etc...)

-    set_tiny_meta_header_in_use(ptr + (msize << SHIFT_TINY_QUANTUM), 1);
-    szone->tiny_bytes_free_at_end = (NUM_TINY_BLOCKS - msize) << SHIFT_TINY_QUANTUM;
+    set_tiny_meta_header_in_use(ptr + TINY_BYTES_FOR_MSIZE(msize), 1);
+    szone->tiny_bytes_free_at_end = TINY_BYTES_FOR_MSIZE(NUM_TINY_BLOCKS - msize);

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In tiny_malloc_from_region_no_lock(), ptr=%p, msize=%d\n", ptr, msize);
+       malloc_printf("in tiny_malloc_from_region_no_lock(), ptr=%p, msize=%d\n", ptr, msize);

     }
 #endif
     return ptr;
 }
 
 static INLINE boolean_t
     }
 #endif
     return ptr;
 }
 
 static INLINE boolean_t

-try_realloc_tiny_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) {
+try_realloc_tiny_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size)
+{

     // returns 1 on success
     // returns 1 on success

-    msize_t    index = (((unsigned)ptr) >> SHIFT_TINY_QUANTUM) & (NUM_TINY_BLOCKS - 1);
-    msize_t    old_msize = old_size >> SHIFT_TINY_QUANTUM;
-    unsigned   next_index = index + old_msize;
-    // malloc_printf("try_realloc_tiny_in_place %p %d %d\n", ptr, old_size, new_size);
+    msize_t    index;
+    msize_t    old_msize;
+    unsigned   next_index;
+    void       *next_block;
+    boolean_t  is_free;
+    msize_t    next_msize, coalesced_msize, leftover_msize;
+    void       *leftover;
+
+    index = TINY_INDEX_FOR_PTR(ptr);
+    old_msize = TINY_MSIZE_FOR_BYTES(old_size);
+    next_index = index + old_msize;
+    

     if (next_index >= NUM_TINY_BLOCKS) {
     if (next_index >= NUM_TINY_BLOCKS) {

-       // malloc_printf("try_realloc_tiny_in_place can't take place at end %p %d %d %d\n", ptr, old_size, new_size, next_index);

        return 0;
     }
        return 0;
     }

-    void       *next_block = (char *)ptr + old_size;
+    next_block = (char *)ptr + old_size;

     SZONE_LOCK(szone);
     SZONE_LOCK(szone);

-    boolean_t  is_free = tiny_meta_header_is_free(next_block);
+    is_free = tiny_meta_header_is_free(next_block);

     if (!is_free) {
        SZONE_UNLOCK(szone);
        return 0; // next_block is in use;
     }
     if (!is_free) {
        SZONE_UNLOCK(szone);
        return 0; // next_block is in use;
     }

-    msize_t    next_msize = TINY_FREE_SIZE(next_block);
-    if (old_size + (next_msize >> SHIFT_TINY_QUANTUM) < new_size) {
-       // malloc_printf("try_realloc_tiny_in_place can't %p too small %d\n", next_block, next_msize);
+    next_msize = get_tiny_free_size(next_block);
+    if (old_size + TINY_MSIZE_FOR_BYTES(next_msize) < new_size) {

        SZONE_UNLOCK(szone);
        return 0; // even with next block, not enough
     }
     tiny_free_list_remove_ptr(szone, next_block, next_msize);
     set_tiny_meta_header_middle(next_block); // clear the meta_header to enable coalescing backwards
        SZONE_UNLOCK(szone);
        return 0; // even with next block, not enough
     }
     tiny_free_list_remove_ptr(szone, next_block, next_msize);
     set_tiny_meta_header_middle(next_block); // clear the meta_header to enable coalescing backwards

-    msize_t    coalesced_msize = (new_size - old_size + TINY_QUANTUM - 1) >> SHIFT_TINY_QUANTUM;
-    msize_t    leftover_msize = next_msize - coalesced_msize;
-    // malloc_printf("Realloc in place for %p;  current size=%d next_msize=%d wanted=%d\n", ptr, old_size, next_msize, new_size);
+    coalesced_msize = TINY_MSIZE_FOR_BYTES(new_size - old_size + TINY_QUANTUM - 1);
+    leftover_msize = next_msize - coalesced_msize;

     if (leftover_msize) {
     if (leftover_msize) {

-       void    *leftover = next_block + (coalesced_msize << SHIFT_TINY_QUANTUM);
-       // malloc_printf("Leftover in realloc in place %p leftover_msize=%d\n", leftover, leftover_msize);
+       leftover = next_block + TINY_BYTES_FOR_MSIZE(coalesced_msize);

        tiny_free_list_add_ptr(szone, leftover, leftover_msize);
     }
     set_tiny_meta_header_in_use(ptr, old_msize + coalesced_msize);
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
        tiny_free_list_add_ptr(szone, leftover, leftover_msize);
     }
     set_tiny_meta_header_in_use(ptr, old_msize + coalesced_msize);
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In try_realloc_tiny_in_place(), ptr=%p, msize=%d\n", ptr, old_msize + coalesced_msize);
+       malloc_printf("in try_realloc_tiny_in_place(), ptr=%p, msize=%d\n", ptr, old_msize + coalesced_msize);

     }
 #endif
     }
 #endif

-    szone->num_bytes_in_tiny_objects += coalesced_msize << SHIFT_TINY_QUANTUM;
+    szone->num_bytes_in_tiny_objects += TINY_BYTES_FOR_MSIZE(coalesced_msize);

     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);

-    // malloc_printf("Extended ptr %p for realloc old=%d desired=%d new=%d leftover=%d\n", ptr, (unsigned)old_size, (unsigned)new_size, (unsigned)szone_size(szone, ptr), leftover_msize << SHIFT_TINY_QUANTUM);

     return 1;
 }
 
 static boolean_t
     return 1;
 }
 
 static boolean_t

-szone_check_tiny_region(szone_t *szone, tiny_region_t *region) {
-    vm_address_t       start = TINY_REGION_ADDRESS(*region);
-    void               *ptr = (void *)start;
-    vm_address_t       region_end = TINY_REGION_END(*region);
-    boolean_t          prev_free = 0;
-    if (region == szone->tiny_regions + szone->num_tiny_regions - 1) region_end -= szone->tiny_bytes_free_at_end;
-    // malloc_printf("szone_check_tiny_region: szone=%p region=%p start=%p ptr=%p region_end=%p\n", szone, region, start, ptr, region_end);
-    while ((vm_address_t)ptr < region_end) {
-       boolean_t       is_free;
-       msize_t         msize = get_tiny_meta_header(ptr, &is_free);
-       if (is_free && !msize && (ptr == (void *)start)) {
-           // the entire region is free
+tiny_check_region(szone_t *szone, region_t region)
+{
+    uintptr_t   start, ptr, region_end;
+    boolean_t   prev_free = 0;
+    boolean_t   is_free;
+    msize_t     msize;
+    free_list_t *free_head;
+    void        *follower, *previous, *next;
+
+    /* establish region limits */
+    start = (uintptr_t)TINY_REGION_ADDRESS(region);
+    ptr = start;
+    region_end = (uintptr_t)TINY_REGION_END(region);
+
+    /*
+     * The last region may have a trailing chunk which has not been converted into inuse/freelist
+     * blocks yet.
+     */
+    if (region == szone->last_tiny_region)
+           region_end -= szone->tiny_bytes_free_at_end;
+
+
+    /*
+     * Scan blocks within the region.
+     */
+    while (ptr < region_end) {
+       /*
+        * If the first block is free, and its size is 65536 (msize = 0) then the entire region is
+        * free.
+        */
+       msize = get_tiny_meta_header((void *)ptr, &is_free);
+       if (is_free && !msize && (ptr == start)) {

            return 1;
        }
            return 1;
        }

-       // malloc_printf("tiny %p [%d %d]\n", ptr, msize, is_free);
-       if (! msize) {
-           malloc_printf("*** malloc[%d]: invariant broken for tiny block %p this msize=%d - size is too small\n", getpid(), ptr, msize);
+
+       /*
+        * If the block's size is 65536 (msize = 0) then since we're not the first entry the size is
+        * corrupt.
+        */
+       if (!msize) {
+           malloc_printf("*** invariant broken for tiny block %p this msize=%d - size is too small\n",
+             ptr, msize);

            return 0;
        }
            return 0;
        }

+

        if (!is_free) {
        if (!is_free) {

-           // this block is in use
+           /*
+            * In use blocks cannot be more than 31 quanta large.
+            */

            prev_free = 0;
            prev_free = 0;

-           if (msize > 31*TINY_QUANTUM) {
-               malloc_printf("*** malloc[%d]: invariant broken for %p this tiny msize=%d[%p] - size is too large\n", getpid(), ptr, msize, msize);
+           if (msize > 31 * TINY_QUANTUM) {
+               malloc_printf("*** invariant broken for %p this tiny msize=%d[%p] - size is too large\n",
+                 ptr, msize, msize);

                return 0;
            }
                return 0;
            }

-           ptr += msize << SHIFT_TINY_QUANTUM;
+           /* move to next block */
+           ptr += TINY_BYTES_FOR_MSIZE(msize);

        } else {
        } else {

-           // free pointer
+           /*
+            * Free blocks must have been coalesced, we cannot have a free block following another
+            * free block.
+            */

            if (prev_free) {
            if (prev_free) {

-               malloc_printf("*** malloc[%d]: invariant broken for free block %p this tiny msize=%d: two free blocks in a row\n", getpid(), ptr, msize);
+               malloc_printf("*** invariant broken for free block %p this tiny msize=%d: two free blocks in a row\n",
+                 ptr, msize);

                return 0;
            }
            prev_free = 1;
                return 0;
            }
            prev_free = 1;

-           free_list_t *free_head = ptr;
+           /*
+            * Check the integrity of this block's entry in its freelist.
+            */
+           free_head = (free_list_t *)ptr;

            free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);
            free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);

-           if (free_head->previous && !tiny_meta_header_is_free(free_head->previous)) {
-               malloc_printf("*** malloc[%d]: invariant broken for %p (previous %p is not a free pointer)\n", getpid(), ptr, free_head->previous);
+           previous = free_list_unchecksum_ptr(free_head->previous);
+           next = free_list_unchecksum_ptr(free_head->next);
+           if (previous && !tiny_meta_header_is_free(previous)) {
+               malloc_printf("*** invariant broken for %p (previous %p is not a free pointer)\n",
+                 ptr, previous);

                return 0;
            }
                return 0;
            }

-           if (free_head->next && !tiny_meta_header_is_free(free_head->next)) {
-               malloc_printf("*** malloc[%d]: invariant broken for %p (next in free list %p is not a free pointer)\n", getpid(), ptr, free_head->next);
+           if (next && !tiny_meta_header_is_free(next)) {
+               malloc_printf("*** invariant broken for %p (next in free list %p is not a free pointer)\n",
+                 ptr, next);

                return 0;
            }
                return 0;
            }

-           void        *follower = FOLLOWING_TINY_PTR(ptr, msize);
-           if ((follower != (void *)region_end) && (TINY_PREVIOUS_MSIZE(follower) != msize)) {
-               malloc_printf("*** malloc[%d]: invariant broken for tiny free %p followed by %p in region [%p-%p] (end marker incorrect) should be %d; in fact %d\n", getpid(), ptr, follower, TINY_REGION_ADDRESS(*region), region_end, msize, TINY_PREVIOUS_MSIZE(follower));
+           /*
+            * Check the free block's trailing size value.
+            */
+           follower = FOLLOWING_TINY_PTR(ptr, msize);
+           if (((uintptr_t)follower != region_end) && (get_tiny_previous_free_msize(follower) != msize)) {
+               malloc_printf("*** invariant broken for tiny free %p followed by %p in region [%p-%p] "
+                 "(end marker incorrect) should be %d; in fact %d\n",
+                 ptr, follower, TINY_REGION_ADDRESS(region), region_end, msize, get_tiny_previous_free_msize(follower));

                return 0;
            }
                return 0;
            }

-           ptr = follower;
+           /* move to next block */
+           ptr = (uintptr_t)follower;

        }
     }
        }
     }

-    if (ptr != (void *)region_end) {
-       malloc_printf("*** malloc[%d]: invariant broken for region end %p - %p\n", getpid(), ptr, region_end);
+    /*
+     * Ensure that we scanned the entire region
+     */
+    if (ptr != region_end) {
+       malloc_printf("*** invariant broken for region end %p - %p\n", ptr, region_end);

        return 0;
     }
        return 0;
     }

-    if (region == szone->tiny_regions + szone->num_tiny_regions - 1) {
+    /*
+     * Check the trailing block's integrity.
+     */
+    if (region == szone->last_tiny_region) {

        if (szone->tiny_bytes_free_at_end) {
        if (szone->tiny_bytes_free_at_end) {

-           boolean_t   is_free;
-           msize_t     msize = get_tiny_meta_header(ptr, &is_free);
+           msize = get_tiny_meta_header((void *)ptr, &is_free);

            if (is_free || (msize != 1)) {
            if (is_free || (msize != 1)) {

-               malloc_printf("*** malloc[%d]: invariant broken for blocker block %p - %d %d\n", getpid(), ptr, msize, is_free);
+               malloc_printf("*** invariant broken for blocker block %p - %d %d\n", ptr, msize, is_free);

            }
        }
     }
            }
        }
     }


@@ -939,197 +1477,260 @@ szone_check_tiny_region(szone_t *szone, tiny_region_t *region) {
 }
 
 static kern_return_t
 }
 
 static kern_return_t

-tiny_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t region_address, unsigned short num_regions, size_t tiny_bytes_free_at_end, memory_reader_t reader, vm_range_recorder_t recorder) {
-    tiny_region_t      *regions;
-    unsigned           index = 0;
-    vm_range_t         buffer[MAX_RECORDER_BUFFER];
-    unsigned           count = 0;
-    kern_return_t      err;
-    err = reader(task, region_address, sizeof(tiny_region_t) * num_regions, (void **)&regions);
-    if (err) return err;
-    while (index < num_regions) {
-       // unsigned             num_in_use = 0;
-       // unsigned             num_free = 0;
-       tiny_region_t   region = regions[index];
-       vm_range_t      range = {TINY_REGION_ADDRESS(region), TINY_REGION_SIZE};
-       // malloc_printf("Enumerating tiny ptrs for tiny region starting at %p\n", range.address);
-       if (type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) {
-           vm_range_t  admin_range = {range.address + (1 << TINY_BLOCKS_ALIGN), range.size - (1 << TINY_BLOCKS_ALIGN)};
-           recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &admin_range, 1);
-       }
-       if (type_mask & (MALLOC_PTR_REGION_RANGE_TYPE | MALLOC_ADMIN_REGION_RANGE_TYPE)) {
-           vm_range_t  ptr_range = {range.address, 1 << TINY_BLOCKS_ALIGN};
-           recorder(task, context, MALLOC_PTR_REGION_RANGE_TYPE, &ptr_range, 1);
-       }
-       if (type_mask & MALLOC_PTR_IN_USE_RANGE_TYPE) {
-           unsigned char       *mapped_region;
-           err = reader(task, range.address, range.size, (void **)&mapped_region);
-           if (err) return err;
-           unsigned char       *block_header = (unsigned char *)(mapped_region + (1 << TINY_BLOCKS_ALIGN));
-           unsigned char       *in_use = block_header + (NUM_TINY_BLOCKS >> 3) + 4;
-           unsigned            block_index = 0;
-           unsigned            block_limit = NUM_TINY_BLOCKS;
-           if (index == num_regions - 1)
-               block_limit -= (tiny_bytes_free_at_end >> SHIFT_TINY_QUANTUM);
-           while (block_index < block_limit) {
-               boolean_t       is_free = ! BITARRAY_BIT(in_use, block_index);
-               msize_t         msize;
-               if (is_free) {
-                   void        *mapped_ptr = mapped_region + (block_index << SHIFT_TINY_QUANTUM);
-                   msize = TINY_FREE_SIZE(mapped_ptr);
-                   // printf("free:  index=%x mapped=%p true_addr=%p msize=%d\n", block_index, mapped_ptr, (void *)range.address + (block_index << SHIFT_TINY_QUANTUM), msize);
-                   // num_free++;
-                   if (!msize) break;
-               } else {
-                   msize = 1; 
-                   unsigned    bit = block_index + 1;
-                   while (! BITARRAY_BIT(block_header, bit)) { bit++; msize ++; }
-                   // printf("in_use: index=%x true_addr=%p msize=%d\n", block_index, (void *)range.address + (block_index << SHIFT_TINY_QUANTUM), msize);
-                   // num_in_use++;
-                   buffer[count].address = range.address + (block_index << SHIFT_TINY_QUANTUM);
-                   buffer[count].size = msize << SHIFT_TINY_QUANTUM;
-                   count++;
-                   if (count >= MAX_RECORDER_BUFFER) {
-                       recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
-                       count = 0;
-                   }
-               }
-               block_index += msize;
-           }
-       }
-       // malloc_printf("Found in tiny region %d in_use and %d free\n", num_in_use, num_free);
-       index++;
-    }
-    if (count) {
-       recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
-    }
-    return 0;
+tiny_in_use_enumerator(task_t task, void *context, unsigned type_mask, szone_t *szone, memory_reader_t reader, vm_range_recorder_t recorder)
+{
+  size_t num_regions = szone->num_tiny_regions_allocated;
+  void *last_tiny_free = szone->last_tiny_free;
+  size_t index;
+  region_t     *regions;
+  vm_range_t           buffer[MAX_RECORDER_BUFFER];
+  unsigned             count = 0;
+  kern_return_t        err;
+  region_t     region;
+  vm_range_t           range;
+  vm_range_t           admin_range;
+  vm_range_t           ptr_range;
+  unsigned char        *mapped_region;
+  unsigned char        *block_header;
+  unsigned char        *in_use;
+  unsigned             block_index;
+  unsigned             block_limit;
+  boolean_t            is_free;
+  msize_t              msize;
+  void         *mapped_ptr;
+  unsigned             bit;
+  vm_address_t last_tiny_free_ptr = 0;
+  msize_t last_tiny_free_msize = 0;
+  
+  if (last_tiny_free) {
+    last_tiny_free_ptr = (uintptr_t) last_tiny_free & ~(TINY_QUANTUM - 1);
+    last_tiny_free_msize = (uintptr_t) last_tiny_free & (TINY_QUANTUM - 1);
+  }
+  
+  err = reader(task, (vm_address_t)szone->tiny_regions, sizeof(region_t) * num_regions, (void **)&regions);
+  if (err) return err;
+  for (index = 0; index < num_regions; ++index) {
+    region = regions[index];
+    if (region) {
+      range.address = (vm_address_t)TINY_REGION_ADDRESS(region);
+      range.size = (vm_size_t)TINY_REGION_SIZE;
+      if (type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) {
+        admin_range.address = range.address + TINY_HEADER_START;
+        admin_range.size = TINY_HEADER_SIZE;
+        recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &admin_range, 1);
+      }
+      if (type_mask & (MALLOC_PTR_REGION_RANGE_TYPE | MALLOC_ADMIN_REGION_RANGE_TYPE)) {
+        ptr_range.address = range.address;
+        ptr_range.size = NUM_TINY_BLOCKS * TINY_QUANTUM;
+        recorder(task, context, MALLOC_PTR_REGION_RANGE_TYPE, &ptr_range, 1);
+      }
+      if (type_mask & MALLOC_PTR_IN_USE_RANGE_TYPE) {
+        err = reader(task, range.address, range.size, (void **)&mapped_region);
+        if (err)
+          return err;
+        
+        block_header = (unsigned char *)(mapped_region + TINY_HEADER_START);
+        in_use = TINY_INUSE_FOR_HEADER(block_header);
+        block_index = 0;
+        block_limit = NUM_TINY_BLOCKS;
+        if (region == szone->last_tiny_region)
+          block_limit -= TINY_MSIZE_FOR_BYTES(szone->tiny_bytes_free_at_end);
+        
+        while (block_index < block_limit) {
+          vm_size_t block_offset = TINY_BYTES_FOR_MSIZE(block_index);
+          is_free = !BITARRAY_BIT(in_use, block_index);
+          if (is_free) {
+            mapped_ptr = mapped_region + block_offset;
+            
+            // mapped_region, the address at which 'range' in 'task' has been
+            // mapped into our process, is not necessarily aligned to 
+            // TINY_BLOCKS_ALIGN.
+            //
+            // Since the code in get_tiny_free_size() assumes the pointer came
+            // from a properly aligned tiny region, and mapped_region is not
+            // necessarily aligned, then do the size calculation directly.
+            // If the next bit is set in the header bitmap, then the size is one
+            // quantum.  Otherwise, read the size field.
+            if (!BITARRAY_BIT(block_header, block_index+1))
+              msize = TINY_FREE_SIZE(mapped_ptr);
+            else
+              msize = 1;
+            
+            if (!msize)
+              break;
+          } else if (range.address + block_offset != last_tiny_free_ptr) {
+            msize = 1; 
+            bit = block_index + 1;
+            while (! BITARRAY_BIT(block_header, bit)) {
+              bit++;
+              msize ++;
+            }
+            buffer[count].address = range.address + block_offset;
+            buffer[count].size = TINY_BYTES_FOR_MSIZE(msize);
+            count++;
+            if (count >= MAX_RECORDER_BUFFER) {
+              recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
+              count = 0;
+            }
+          } else {
+            // Block is not free but it matches last_tiny_free_ptr so even
+            // though it is not marked free in the bitmap, we treat it as if
+            // it is and move on
+            msize = last_tiny_free_msize;
+          }
+          block_index += msize;
+        }
+      }
+    }
+  }
+  if (count) {
+    recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
+  }
+  return 0;

 }
 
 }
 

-static INLINE void *
-tiny_malloc_from_free_list(szone_t *szone, msize_t msize) {
+static void *
+tiny_malloc_from_free_list(szone_t *szone, msize_t msize)
+{

     // Assumes we've locked the region
     // Assumes we've locked the region

-    void       *ptr;
-    msize_t    this_msize;
-    grain_t    slot = msize-1;
-    free_list_t        **free_list = szone->tiny_free_list;
-    free_list_t        **the_slot = free_list + slot;
+    free_list_t                *ptr;
+    msize_t         this_msize;
+    grain_t         slot = msize - 1;
+    free_list_t                **free_list = szone->tiny_free_list;
+    free_list_t                **the_slot = free_list + slot;
+    free_list_t                *next;
+    free_list_t                **limit;
+    unsigned           bitmap;
+    msize_t         leftover_msize;
+    free_list_t                *leftover_ptr;
+
+    // Assumes locked
+    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
+    
+    // Look for an exact match by checking the freelist for this msize.
+    // 

     ptr = *the_slot;
     if (ptr) {
     ptr = *the_slot;
     if (ptr) {

-       free_list_t     *next;
-       next = ((free_list_t *)ptr)->next;
-       if (next) {
-           next->previous = NULL;
-           free_list_set_checksum(szone, next);
-       }
-       *the_slot = next;
-       this_msize = msize;
+        next = free_list_unchecksum_ptr(ptr->next);
+        if (next) {
+            next->previous = ptr->previous;
+        } else {
+            BITMAP32_CLR(szone->tiny_bitmap, slot);
+        }
+        *the_slot = next;
+        this_msize = msize;

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone,ptr)) {
-           malloc_printf("In tiny_malloc_from_free_list(), exact match ptr=%p, this_msize=%d\n", ptr, this_msize);
-       }
+        if (LOG(szone, ptr)) {
+            malloc_printf("in tiny_malloc_from_free_list(), exact match ptr=%p, this_msize=%d\n", ptr, this_msize);
+        }

 #endif
 #endif

-       goto return_tiny_alloc;
-    }
-    // adjust slot based on bitmap
-    unsigned   bitmap = szone->tiny_bitmap & ~ ((1 << slot) - 1);
-    if (! bitmap) goto try_tiny_malloc_from_end;
-    slot = BITMAP32_FFS(bitmap) - 1;
-    free_list_t        **limit = free_list + NUM_TINY_SLOTS - 1;
+        goto return_tiny_alloc;
+    }
+
+    // Mask off the bits representing slots holding free blocks smaller than the
+    // size we need.  If there are no larger free blocks, try allocating from
+    // the free space at the end of the tiny region.
+    bitmap = szone->tiny_bitmap & ~ ((1 << slot) - 1);
+    if (!bitmap)
+        goto try_tiny_malloc_from_end;
+        
+    slot = BITMAP32_CTZ(bitmap);
+    limit = free_list + NUM_TINY_SLOTS - 1;

     free_list += slot;
     free_list += slot;

+    
+    // Iterate over freelists looking for free blocks, starting at first list
+    // which is not empty, and contains blocks which are large enough to satisfy
+    // our request.

     while (free_list < limit) {
     while (free_list < limit) {

-       // try bigger grains
-       ptr = *free_list;
-       if (ptr) {
-           free_list_t *next;
-           next = ((free_list_t *)ptr)->next;
-           if (next) {
-               next->previous = NULL;
-               free_list_set_checksum(szone, next);
-           }
-           *free_list = next;
-           this_msize = TINY_FREE_SIZE(ptr);
-#if DEBUG_MALLOC
-           if (LOG(szone,ptr)) {
-               malloc_printf("In tiny_malloc_from_free_list(), bigger grain ptr=%p, msize=%d this_msize=%d\n", ptr, msize, this_msize);
-           }
-#endif
-           goto add_leftover_and_proceed;
-       }
-       free_list++;
-    }
-    // we are now looking at the last slot (31)
+        ptr = *free_list;
+        if (ptr) {
+            next = free_list_unchecksum_ptr(ptr->next);
+            *free_list = next;
+            this_msize = get_tiny_free_size(ptr);
+            if (next) {
+                next->previous = ptr->previous;
+            } else {
+                BITMAP32_CLR(szone->tiny_bitmap, this_msize - 1);
+            }
+            goto add_leftover_and_proceed;
+        }
+        free_list++;
+    }
+
+    // We are now looking at the last slot, which contains blocks equal to, or
+    // due to coalescing of free blocks, larger than 31 * tiny quantum size.
+    // If the last freelist is not empty, and the head contains a block that is
+    // larger than our request, then the remainder is put back on the free list.

     ptr = *limit;
     if (ptr) {
     ptr = *limit;
     if (ptr) {

-       free_list_t     *next;
-       this_msize = TINY_FREE_SIZE(ptr);
-       next = ((free_list_t *)ptr)->next;
-       if (this_msize - msize >= NUM_TINY_SLOTS) {
-           // the leftover will go back to the free list, so we optimize by modifying the free list rather than removing the head and then adding back
-           // malloc_printf("Allocation from largest tiny slot %p optimized\n", ptr);
-           msize_t     leftover_msize = this_msize - msize;
-           void        *leftover_ptr = ptr + (msize << SHIFT_TINY_QUANTUM);
-           *limit = leftover_ptr;
-           if (next) {
-               next->previous = leftover_ptr;
-               free_list_set_checksum(szone, next);
-           }
-           ((free_list_t *)leftover_ptr)->next = next;
-           ((free_list_t *)leftover_ptr)->previous = NULL;
-           free_list_set_checksum(szone, leftover_ptr);
-           set_tiny_meta_header_free(leftover_ptr, leftover_msize);
+        free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
+        this_msize = get_tiny_free_size(ptr);
+        next = free_list_unchecksum_ptr(ptr->next);
+        if (this_msize - msize >= NUM_TINY_SLOTS) {
+            // the leftover will go back to the free list, so we optimize by
+            // modifying the free list rather than a pop and push of the head
+            leftover_msize = this_msize - msize;
+            leftover_ptr = (free_list_t *)((unsigned char *)ptr + TINY_BYTES_FOR_MSIZE(msize));
+            *limit = leftover_ptr;
+            if (next) {
+                next->previous.u = free_list_checksum_ptr(leftover_ptr);
+            }
+            leftover_ptr->previous = ptr->previous;
+            leftover_ptr->next = ptr->next;
+            set_tiny_meta_header_free(leftover_ptr, leftover_msize);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-           if (LOG(szone,ptr)) {
-               malloc_printf("In tiny_malloc_from_free_list(), last slot ptr=%p, msize=%d this_msize=%d\n", ptr, msize, this_msize);
-           }
+            if (LOG(szone,ptr)) {
+            malloc_printf("in tiny_malloc_from_free_list(), last slot ptr=%p, msize=%d this_msize=%d\n", ptr, msize, this_msize);
+            }

 #endif
 #endif

-           this_msize = msize;
-           goto return_tiny_alloc;
-       }
-       *limit = next;
-       if (next) {
-           next->previous = NULL;
-           free_list_set_checksum(szone, next);
-       }
-       goto add_leftover_and_proceed;
+            this_msize = msize;
+            goto return_tiny_alloc;
+        }
+        if (next) {
+            next->previous = ptr->previous;
+        }
+        *limit = next;
+        goto add_leftover_and_proceed;

     }
     }

+

 try_tiny_malloc_from_end:
     // Let's see if we can use szone->tiny_bytes_free_at_end
 try_tiny_malloc_from_end:
     // Let's see if we can use szone->tiny_bytes_free_at_end

-    if (szone->tiny_bytes_free_at_end >= (msize << SHIFT_TINY_QUANTUM)) {
-       ptr = (void *)(TINY_REGION_END(szone->tiny_regions[szone->num_tiny_regions-1]) - szone->tiny_bytes_free_at_end);
-       szone->tiny_bytes_free_at_end -= msize << SHIFT_TINY_QUANTUM;
-       if (szone->tiny_bytes_free_at_end) {
-           // let's add an in use block after ptr to serve as boundary
-           set_tiny_meta_header_in_use(ptr + (msize << SHIFT_TINY_QUANTUM), 1);
-       }
-       this_msize = msize;
+    if (szone->tiny_bytes_free_at_end >= TINY_BYTES_FOR_MSIZE(msize)) {
+        ptr = (free_list_t *)(TINY_REGION_END(szone->last_tiny_region) - szone->tiny_bytes_free_at_end);
+        szone->tiny_bytes_free_at_end -= TINY_BYTES_FOR_MSIZE(msize);
+        if (szone->tiny_bytes_free_at_end) {
+            // let's add an in use block after ptr to serve as boundary
+            set_tiny_meta_header_in_use((unsigned char *)ptr + TINY_BYTES_FOR_MSIZE(msize), 1);
+        }
+        this_msize = msize;

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone,ptr)) {
-           malloc_printf("In tiny_malloc_from_free_list(), from end ptr=%p, msize=%d\n", ptr, msize);
-       }
+        if (LOG(szone, ptr)) {
+            malloc_printf("in tiny_malloc_from_free_list(), from end ptr=%p, msize=%d\n", ptr, msize);
+        }

 #endif
 #endif

-       goto return_tiny_alloc;
+        goto return_tiny_alloc;

     }
     return NULL;
     }
     return NULL;

+

 add_leftover_and_proceed:
 add_leftover_and_proceed:

-    // malloc_printf("For msize=%d found tiny in free_list (slot=%d) this_msize=%d\n", msize, free_list - szone->tiny_free_list, this_msize);

     if (!this_msize || (this_msize > msize)) {
     if (!this_msize || (this_msize > msize)) {

-       msize_t leftover_msize = this_msize - msize;
-       void    *leftover_ptr = ptr + (msize << SHIFT_TINY_QUANTUM);
+        leftover_msize = this_msize - msize;
+        leftover_ptr = (free_list_t *)((unsigned char *)ptr + TINY_BYTES_FOR_MSIZE(msize));

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone,ptr)) {
-           malloc_printf("In tiny_malloc_from_free_list(), adding leftover ptr=%p, this_msize=%d\n", ptr, this_msize);
-       }
+        if (LOG(szone,ptr)) {
+            malloc_printf("in tiny_malloc_from_free_list(), adding leftover ptr=%p, this_msize=%d\n", ptr, this_msize);
+        }

 #endif
 #endif

-       tiny_free_list_add_ptr(szone, leftover_ptr, leftover_msize);
-       this_msize = msize;
+        tiny_free_list_add_ptr(szone, leftover_ptr, leftover_msize);
+        this_msize = msize;

     }
     }

+    

 return_tiny_alloc:
     szone->num_tiny_objects++;
 return_tiny_alloc:
     szone->num_tiny_objects++;

-    szone->num_bytes_in_tiny_objects += this_msize << SHIFT_TINY_QUANTUM;
+    szone->num_bytes_in_tiny_objects += TINY_BYTES_FOR_MSIZE(this_msize);

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In tiny_malloc_from_free_list(), ptr=%p, this_msize=%d, msize=%d\n", ptr, this_msize, msize);
+       malloc_printf("in tiny_malloc_from_free_list(), ptr=%p, this_msize=%d, msize=%d\n", ptr, this_msize, msize);

     }
 #endif
     set_tiny_meta_header_in_use(ptr, this_msize);
     }
 #endif
     set_tiny_meta_header_in_use(ptr, this_msize);


@@ -1137,52 +1738,51 @@ return_tiny_alloc:
 }
 
 static INLINE void *
 }
 
 static INLINE void *

-tiny_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested) {
+tiny_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested)
+{

     boolean_t  locked = 0;
     void       *ptr;
     boolean_t  locked = 0;
     void       *ptr;

+

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-    if (! msize) {
-       szone_error(szone, "Invariant broken (!msize) in allocation (region)", NULL);
+    if (!msize) {
+       szone_error(szone, "invariant broken (!msize) in allocation (region)", NULL, NULL);
+       return(NULL);

     }
 #endif
 #if TINY_CACHE
     }
 #endif
 #if TINY_CACHE

-    ptr = (void *)szone->last_tiny_free;
-    if ((((unsigned)ptr) & (TINY_QUANTUM - 1)) == msize) {
+    ptr = szone->last_tiny_free;
+    if ((((uintptr_t)ptr) & (TINY_QUANTUM - 1)) == msize) {

        // we have a candidate - let's lock to make sure
        LOCK_AND_NOTE_LOCKED(szone, locked);
        // we have a candidate - let's lock to make sure
        LOCK_AND_NOTE_LOCKED(szone, locked);

-       if (ptr == (void *)szone->last_tiny_free) {
+       if (ptr == szone->last_tiny_free) {

            szone->last_tiny_free = NULL;
            szone->last_tiny_free = NULL;

-           // malloc_printf("using last_tiny_free\n");

            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);
            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);

-           ptr = (void *)((unsigned)ptr & ~ (TINY_QUANTUM - 1));
+           ptr = (void *)((uintptr_t)ptr & ~ (TINY_QUANTUM - 1));

            if (cleared_requested) {
            if (cleared_requested) {

-               memset(ptr, 0, msize << SHIFT_TINY_QUANTUM);
+               memset(ptr, 0, TINY_BYTES_FOR_MSIZE(msize));

            }
 #if DEBUG_MALLOC
            if (LOG(szone,ptr)) {
            }
 #if DEBUG_MALLOC
            if (LOG(szone,ptr)) {

-               malloc_printf("In tiny_malloc_should_clear(), tiny cache ptr=%p, msize=%d\n", ptr, msize);
+               malloc_printf("in tiny_malloc_should_clear(), tiny cache ptr=%p, msize=%d\n", ptr, msize);

            }
 #endif
            return ptr;
        }
            }
 #endif
            return ptr;
        }

-       // malloc_printf("optimistic locking for last_tiny_free failed\n");

     }
 #endif
     // Except in rare occasions where we need to add a new region, we are going to end up locking, so we might as well lock right away to avoid doing unnecessary optimistic probes
     if (!locked) LOCK_AND_NOTE_LOCKED(szone, locked);
     ptr = tiny_malloc_from_free_list(szone, msize);
     }
 #endif
     // Except in rare occasions where we need to add a new region, we are going to end up locking, so we might as well lock right away to avoid doing unnecessary optimistic probes
     if (!locked) LOCK_AND_NOTE_LOCKED(szone, locked);
     ptr = tiny_malloc_from_free_list(szone, msize);

-    // malloc_printf("tiny_malloc_from_free_list(%d) returned %p\n", msize, ptr);

     if (ptr) {
        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        if (cleared_requested) {
     if (ptr) {
        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        if (cleared_requested) {

-           memset(ptr, 0, msize << SHIFT_TINY_QUANTUM);
+           memset(ptr, 0, TINY_BYTES_FOR_MSIZE(msize));

        }
        return ptr;
     }
     ptr = tiny_malloc_from_region_no_lock(szone, msize);
        }
        return ptr;
     }
     ptr = tiny_malloc_from_region_no_lock(szone, msize);

-    // malloc_printf("tiny_malloc_from_region_no_lock returned %p for msize=%d\n", ptr, msize);

     // we don't clear because this freshly allocated space is pristine
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
     // we don't clear because this freshly allocated space is pristine
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);


@@ -1190,43 +1790,50 @@ tiny_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_reques
 }
 
 static INLINE void
 }
 
 static INLINE void

-free_tiny(szone_t *szone, void *ptr, tiny_region_t *tiny_region) {
+free_tiny(szone_t *szone, void *ptr, region_t *tiny_region)
+{
+    msize_t    msize;
+    boolean_t  is_free;
+#if TINY_CACHE
+    void *ptr2;
+#endif
+

     // ptr is known to be in tiny_region
     SZONE_LOCK(szone);
 #if TINY_CACHE
     // ptr is known to be in tiny_region
     SZONE_LOCK(szone);
 #if TINY_CACHE

-    void *ptr2 = szone->last_tiny_free;
-    if (ptr == (void *)((unsigned)ptr2 & ~ (TINY_QUANTUM - 1))) {
-       szone_error(szone, "Double free", ptr);
+    ptr2 = szone->last_tiny_free;
+    /* check that we don't already have this pointer in the cache */
+    if (ptr == (void *)((uintptr_t)ptr2 & ~ (TINY_QUANTUM - 1))) {
+       szone_error(szone, "double free", ptr, NULL);

        return;
     }
 #endif /* TINY_CACHE */
        return;
     }
 #endif /* TINY_CACHE */

-    boolean_t  is_free;
-    msize_t    msize = get_tiny_meta_header(ptr, &is_free);
+    msize = get_tiny_meta_header(ptr, &is_free);

     if (is_free) {
     if (is_free) {

-       szone_error(szone, "Double free", ptr);
+       szone_error(szone, "double free", ptr, NULL);

        return;
     }
        return;
     }

-    // malloc_printf("%p[%x]\n", ptr, msize);

 #if DEBUG_MALLOC
     if (!msize) {
        malloc_printf("*** szone_free() block in use is too large: %p\n", ptr);
 #if DEBUG_MALLOC
     if (!msize) {
        malloc_printf("*** szone_free() block in use is too large: %p\n", ptr);

+       return;

     }
 #endif
 #if TINY_CACHE
     if (msize < TINY_QUANTUM) {        // to see if the bits fit in the last 4 bits
     }
 #endif
 #if TINY_CACHE
     if (msize < TINY_QUANTUM) {        // to see if the bits fit in the last 4 bits

-       szone->last_tiny_free = (void *)(((unsigned)ptr) | msize);
+    if ((szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) && msize)
+        memset(ptr, 0x55, TINY_BYTES_FOR_MSIZE(msize));
+    szone->last_tiny_free = (void *)(((uintptr_t)ptr) | msize);

        if (!ptr2) {
        if (!ptr2) {

-           // malloc_printf("stuffing last_tiny_free\n");

            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);
            return;
        }
            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);
            return;
        }

-       // malloc_printf("replacing previous last_tiny_free %p with %p\n", ptr2, szone->last_tiny_free);
-       msize = (unsigned)ptr2 & (TINY_QUANTUM - 1);
-       ptr = (void *)(((unsigned)ptr2) & ~ (TINY_QUANTUM - 1));
+       msize = (uintptr_t)ptr2 & (TINY_QUANTUM - 1);
+       ptr = (void *)(((uintptr_t)ptr2) & ~(TINY_QUANTUM - 1));

        tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
        if (!tiny_region) {
        tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
        if (!tiny_region) {

-           szone_error(szone, "Double free (tiny cache)", ptr);
+           szone_error(szone, "double free (tiny cache)", ptr, NULL);

        }
     }
 #endif
        }
     }
 #endif


@@ -1236,260 +1843,267 @@ free_tiny(szone_t *szone, void *ptr, tiny_region_t *tiny_region) {
 }
 
 static void
 }
 
 static void

-print_tiny_free_list(szone_t *szone) {
-    grain_t            slot = 0;
-    malloc_printf("Tiny free sizes: ");
-    while (slot < NUM_TINY_SLOTS) {
-       free_list_t     *ptr = szone->tiny_free_list[slot];
-       if (ptr) {
-           malloc_printf("%s%y[%d]; ", (slot == NUM_TINY_SLOTS-1) ? ">=" : "", (slot+1)*TINY_QUANTUM, free_list_count(ptr));
+print_tiny_free_list(szone_t *szone)
+{
+    grain_t    slot = 0;
+    free_list_t        *ptr;
+    _SIMPLE_STRING b = _simple_salloc();
+
+    if (b) {
+       _simple_sappend(b, "tiny free sizes: ");
+       while (slot < NUM_TINY_SLOTS) {
+           ptr = szone->tiny_free_list[slot];
+           if (ptr) {
+               _simple_sprintf(b, "%s%y[%d]; ", (slot == NUM_TINY_SLOTS-1) ? ">=" : "", (slot+1)*TINY_QUANTUM, free_list_count(ptr));
+           }
+           slot++;

        }
        }

-       slot++;
+       _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX, "%s\n", _simple_string(b));
+       _simple_sfree(b);

     }
     }

-    malloc_printf("\n");

 }
 
 static void
 }
 
 static void

-print_tiny_region(boolean_t verbose, tiny_region_t region, size_t bytes_at_end) {
+print_tiny_region(boolean_t verbose, region_t region, size_t bytes_at_end)
+{

     unsigned   counts[1024];
     unsigned   in_use = 0;
     unsigned   counts[1024];
     unsigned   in_use = 0;

-    vm_address_t       start = TINY_REGION_ADDRESS(region);
-    vm_address_t       current = start;
-    vm_address_t       limit = TINY_REGION_END(region) - bytes_at_end;
+    uintptr_t  start = (uintptr_t)TINY_REGION_ADDRESS(region);
+    uintptr_t  current = start;
+    uintptr_t  limit =  (uintptr_t)TINY_REGION_END(region) - bytes_at_end;
+    boolean_t  is_free;
+    msize_t    msize;
+    unsigned   ci;
+    _SIMPLE_STRING b;
+    

     memset(counts, 0, 1024 * sizeof(unsigned));
     while (current < limit) {
     memset(counts, 0, 1024 * sizeof(unsigned));
     while (current < limit) {

-       boolean_t       is_free;
-       msize_t         msize = get_tiny_meta_header((void *)current, &is_free);
-       // malloc_printf("%p [%d %d]; ", current, msize, is_free);
+       msize = get_tiny_meta_header((void *)current, &is_free);

        if (is_free & !msize && (current == start)) {
            // first block is all free
            break;
        }
        if (!msize) {
        if (is_free & !msize && (current == start)) {
            // first block is all free
            break;
        }
        if (!msize) {

-           malloc_printf("*** Error with %p: msize=%d\n", current, msize);
+           malloc_printf("*** error with %p: msize=%d\n", (void *)current, (unsigned)msize);

            break;
        }
            break;
        }

-       if (! is_free) {
+       if (!is_free) {

            // block in use
            // block in use

-           if (msize > 32) malloc_printf("*** Error at %p msize for in_use is %d\n", current, msize);
-           if (msize < 1024) counts[msize]++;
+           if (msize > 32)
+               malloc_printf("*** error at %p msize for in_use is %d\n", (void *)current, msize);
+           if (msize < 1024)
+               counts[msize]++;

            in_use++;
        }
            in_use++;
        }

-       current += msize << SHIFT_TINY_QUANTUM;
-    }
-    malloc_printf("Tiny region [%p-%p, %y]\t", start, TINY_REGION_END(region), (int)TINY_REGION_SIZE);
-    malloc_printf("In_use=%d ", in_use);
-    if (bytes_at_end) malloc_printf("Untouched=%y ", bytes_at_end);
-    if (verbose && in_use) {
-       unsigned        ci = 0;
-       malloc_printf("\n\tSizes in use: "); 
-       while (ci < 1024) {
-           if (counts[ci]) {
-               malloc_printf("%d[%d] ", ci << SHIFT_TINY_QUANTUM, counts[ci]);
-           }
-           ci++;
+       current += TINY_BYTES_FOR_MSIZE(msize);
+    }
+    if ((b = _simple_salloc()) != NULL) {
+       _simple_sprintf(b, "Tiny region [%p-%p, %y]\t", (void *)start, TINY_REGION_END(region), (int)TINY_REGION_SIZE);
+       _simple_sprintf(b, "In_use=%d ", in_use);
+       if (bytes_at_end) _simple_sprintf(b, "untouched=%ly", bytes_at_end);
+       if (verbose && in_use) {
+           _simple_sappend(b, "\n\tSizes in use: ");
+           for (ci = 0; ci < 1024; ci++)
+               if (counts[ci])
+                   _simple_sprintf(b, "%d[%d]", TINY_BYTES_FOR_MSIZE(ci), counts[ci]);

        }
        }

+       _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX, "%s\n", _simple_string(b));
+       _simple_sfree(b);

     }
     }

-    malloc_printf("\n");

 }
 
 static boolean_t
 }
 
 static boolean_t

-tiny_free_list_check(szone_t *szone, grain_t slot) {
-    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
+tiny_free_list_check(szone_t *szone, grain_t slot)
+{

     unsigned   count = 0;
     free_list_t        *ptr = szone->tiny_free_list[slot];
     free_list_t        *previous = NULL;
     unsigned   count = 0;
     free_list_t        *ptr = szone->tiny_free_list[slot];
     free_list_t        *previous = NULL;

+    boolean_t  is_free;
+
+    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);

     while (ptr) {
     while (ptr) {

-       free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
-       boolean_t       is_free = tiny_meta_header_is_free(ptr);
-       if (! is_free) {
-           malloc_printf("*** malloc[%d]: In-use ptr in free list slot=%d count=%d ptr=%p\n", getpid(), slot, count, ptr);
-           return 0;
-       }
-       if (((unsigned)ptr) & (TINY_QUANTUM - 1)) {
-           malloc_printf("*** malloc[%d]: Unaligned ptr in free list slot=%d  count=%d ptr=%p\n", getpid(), slot, count, ptr);
-           return 0;
-       }
-       if (!tiny_region_for_ptr_no_lock(szone, ptr)) {
-           malloc_printf("*** malloc[%d]: Ptr not in szone slot=%d  count=%d ptr=%p\n", getpid(), slot, count, ptr);
-           return 0;
-       }
-       if (ptr->previous != previous) {
-           malloc_printf("*** malloc[%d]: Previous incorrectly set slot=%d  count=%d ptr=%p\n", getpid(), slot, count, ptr);
-           return 0;
-       }
-       previous = ptr;
-       ptr = ptr->next;
-       count++;
+        free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
+        is_free = tiny_meta_header_is_free(ptr);
+        if (! is_free) {
+            malloc_printf("*** in-use ptr in free list slot=%d count=%d ptr=%p\n", slot, count, ptr);
+            return 0;
+        }
+        if (((uintptr_t)ptr) & (TINY_QUANTUM - 1)) {
+            malloc_printf("*** unaligned ptr in free list slot=%d  count=%d ptr=%p\n", slot, count, ptr);
+            return 0;
+        }
+        if (!tiny_region_for_ptr_no_lock(szone, ptr)) {
+            malloc_printf("*** ptr not in szone slot=%d  count=%d ptr=%p\n", slot, count, ptr);
+            return 0;
+        }
+        if (free_list_unchecksum_ptr(ptr->previous) != previous) {
+            malloc_printf("*** previous incorrectly set slot=%d  count=%d ptr=%p\n", slot, count, ptr);
+            return 0;
+        }
+        previous = ptr;
+        ptr = free_list_unchecksum_ptr(ptr->next);
+        count++;

     }
     }

-    // malloc_printf("tiny_free_list_check passed\n");

     return 1;
 }
 
 /********************* SMALL FREE LIST UTILITIES       ************************/
 
     return 1;
 }
 
 /********************* SMALL FREE LIST UTILITIES       ************************/
 

-static INLINE msize_t *
-small_meta_headers(const void *ptr) {
-    // returns address of meta info
-    unsigned short     shifted_base = ((unsigned)ptr) >> SMALL_BLOCKS_ALIGN;
-    unsigned           headers_start = (shifted_base + 1) << SMALL_BLOCKS_ALIGN;
-    return (msize_t *)headers_start;
-}
-
-static INLINE msize_t
-small_meta_index(const void *ptr) {
-    // returns address of meta info
-    return (((unsigned)ptr) >> SHIFT_SMALL_QUANTUM) & (NUM_SMALL_BLOCKS - 1);
-}
-
-static INLINE msize_t *
-small_meta_header(const void *ptr) {
-    // returns address of meta info
-    msize_t    *meta_headers = small_meta_headers(ptr);
-    msize_t    index = small_meta_index(ptr);
-    return meta_headers + index;
-}
-
+/*
+ * Mark a block as free.  Only the first quantum of a block is marked thusly,
+ * the remainder are marked "middle".
+ */

 static INLINE void
 static INLINE void

-small_meta_header_set_is_free(msize_t *meta_headers, msize_t index, msize_t msize) {
-    // Indicates that the meta_header for index says 'is free'
+small_meta_header_set_is_free(msize_t *meta_headers, unsigned index, msize_t msize)
+{

     meta_headers[index] = msize | SMALL_IS_FREE;
 }
 
     meta_headers[index] = msize | SMALL_IS_FREE;
 }
 

+/*
+ * Mark a block as in use.  Only the first quantum of a block is marked thusly,
+ * the remainder are marked "middle".
+ */

 static INLINE void
 static INLINE void

-small_meta_header_set_in_use(msize_t *meta_headers, msize_t index, msize_t msize) {
-    // Indicates that the meta_header for index says 'in use'
+small_meta_header_set_in_use(msize_t *meta_headers, msize_t index, msize_t msize)
+{

     meta_headers[index] = msize;
 }
 
     meta_headers[index] = msize;
 }
 

+/*
+ * Mark a quantum as being the second or later in a block.
+ */

 static INLINE void
 static INLINE void

-small_meta_header_set_middle(msize_t *meta_headers, msize_t index) {
-    // Indicates that the meta_header for index says 'in the middle of a block'
+small_meta_header_set_middle(msize_t *meta_headers, msize_t index)
+{

     meta_headers[index] = 0;
 }
 
     meta_headers[index] = 0;
 }
 

+// Adds an item to the proper free list
+// Also marks the header of the block properly
+// Assumes szone has been locked    

 static void
 static void

-small_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize) {
-    // Adds an item to the proper free list
-    // Also marks the header of the block properly
-    // Assumes szone has been locked
-    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
-    grain_t    grain = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;
+small_free_list_add_ptr(szone_t *szone, void *ptr, msize_t msize)
+{
+    grain_t    slot = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;

     free_list_t        *free_ptr = ptr;
     free_list_t        *free_ptr = ptr;

-    free_list_t        *free_head = szone->small_free_list[grain];
+    free_list_t        *free_head = szone->small_free_list[slot];
+    void       *follower;
+

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In small_free_list_add_ptr(), ptr=%p, msize=%d\n", ptr, msize);
+        malloc_printf("in %s, ptr=%p, msize=%d\n", __FUNCTION__, ptr, msize);

     }
     }

-    if (((unsigned)ptr) & (SMALL_QUANTUM - 1)) {
-       szone_error(szone, "small_free_list_add_ptr: Unaligned ptr", ptr);
+    if (((uintptr_t)ptr) & (SMALL_QUANTUM - 1)) {
+        szone_error(szone, "small_free_list_add_ptr: Unaligned ptr", ptr, NULL);

     }
 #endif
     if (free_head) {
     }
 #endif
     if (free_head) {

-       free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);
+        free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (free_head->previous) {
-           malloc_printf("ptr=%p grain=%d free_head=%p previous=%p\n", ptr, grain, free_head, free_head->previous);
-           szone_error(szone, "small_free_list_add_ptr: Internal invariant broken (free_head->previous)", ptr);
-       }
-       if (!(small_meta_header(free_head)[0] & SMALL_IS_FREE)) {
-           malloc_printf("ptr=%p grain=%d free_head=%p\n", ptr, grain, free_head);
-           szone_error(szone, "small_free_list_add_ptr: Internal invariant broken (free_head is not a free pointer)", ptr);
-       }
+        if (free_list_unchecksum_ptr(free_head->previous)) {
+            szone_error(szone, "small_free_list_add_ptr: Internal invariant broken (free_head->previous)", ptr,
+               "ptr=%p slot=%d free_head=%p previous=%p\n", ptr, slot, free_head, free_head->previous.p);
+        }
+        if (!SMALL_PTR_IS_FREE(free_head)) {
+            szone_error(szone, "small_free_list_add_ptr: Internal invariant broken (free_head is not a free pointer)", ptr,
+               "ptr=%p slot=%d free_head=%p\n", ptr, slot, free_head);
+        }

 #endif
 #endif

-       free_head->previous = free_ptr;
-       free_list_set_checksum(szone, free_head);
+        free_head->previous.u = free_list_checksum_ptr(free_ptr);

     } else {
     } else {

-       BITMAP32_SET(szone->small_bitmap, grain);
+        BITMAP32_SET(szone->small_bitmap, slot);

     }
     }

-    free_ptr->previous = NULL;
-    free_ptr->next = free_head;
+    free_ptr->previous.p = NULL;
+    free_ptr->next.p = free_head;

     free_list_set_checksum(szone, free_ptr);
     free_list_set_checksum(szone, free_ptr);

-    szone->small_free_list[grain] = free_ptr;
-    void       *follower = ptr + (msize << SHIFT_SMALL_QUANTUM);
+    szone->small_free_list[slot] = free_ptr;
+    follower = ptr + SMALL_BYTES_FOR_MSIZE(msize);

     SMALL_PREVIOUS_MSIZE(follower) = msize;
 }
 
     SMALL_PREVIOUS_MSIZE(follower) = msize;
 }
 

+// Removes item in the proper free list
+// msize could be read, but all callers have it so we pass it in
+// Assumes szone has been locked

 static void
 static void

-small_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize) {
-    // Removes item in the proper free list
-    // msize could be read, but all callers have it so we pass it in
-    // Assumes szone has been locked
-    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
-    grain_t    grain = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;
-    free_list_t        *free_ptr = ptr;
-    free_list_t        *next = free_ptr->next;
-    free_list_t        *previous = free_ptr->previous;
+small_free_list_remove_ptr(szone_t *szone, void *ptr, msize_t msize)
+{
+    grain_t    slot = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;
+    free_list_t        *free_ptr = ptr, *next, *previous;
+    free_list_checksum(szone, free_ptr, __PRETTY_FUNCTION__);
+
+    next = free_list_unchecksum_ptr(free_ptr->next);
+    previous = free_list_unchecksum_ptr(free_ptr->previous);
+

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In small_free_list_remove_ptr(), ptr=%p, msize=%d\n", ptr, msize);
+        malloc_printf("In %s, ptr=%p, msize=%d\n", __FUNCTION__, ptr, msize);

     }
 #endif
     }
 #endif

-    free_list_checksum(szone, free_ptr, __PRETTY_FUNCTION__);
-    if (!previous) {
+    if (!previous) { 
+    // The block to remove is the head of the free list

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (szone->small_free_list[grain] != ptr) {
-           malloc_printf("ptr=%p grain=%d msize=%d szone->small_free_list[grain]=%p\n", ptr, grain, msize, szone->small_free_list[grain]);
-           szone_error(szone, "small_free_list_remove_ptr: Internal invariant broken (szone->small_free_list[grain])", ptr);
-           return;
-       }
+        if (szone->small_free_list[slot] != ptr) {
+            szone_error(szone, "small_free_list_remove_ptr: Internal invariant broken (szone->small_free_list[grain])", ptr,
+              "ptr=%p slot=%d msize=%d szone->small_free_list[slot]=%p\n",
+              ptr, slot, msize, szone->small_free_list[slot]);
+            return;
+        }

 #endif
 #endif

-       szone->small_free_list[grain] = next;
-       if (!next) BITMAP32_CLR(szone->small_bitmap, grain);
+        szone->small_free_list[slot] = next;
+        if (!next) BITMAP32_CLR(szone->small_bitmap, slot);

     } else {
     } else {

-       previous->next = next;
-       free_list_set_checksum(szone, previous);
+        // We know free_ptr is already checksummed, so we don't need to do it
+        // again.
+        previous->next = free_ptr->next;

     }
     if (next) {
     }
     if (next) {

-       next->previous = previous;
-       free_list_set_checksum(szone, next);
+        // We know free_ptr is already checksummed, so we don't need to do it
+        // again.
+        next->previous = free_ptr->previous;

     }
 }
 
     }
 }
 

-static INLINE small_region_t *
-small_region_for_ptr_no_lock(szone_t *szone, const void *ptr) {
-    small_region_t     *region = szone->small_regions;
-    unsigned           num_regions = szone->num_small_regions;
-    unsigned           ptr_shifted = ((unsigned)ptr) >> SMALL_BLOCKS_ALIGN;
-    while (num_regions--) {
-       small_region_t  this = *region;
-       if (ptr_shifted == this) return region;
-       region++;
-    }
-    return NULL;
+static INLINE region_t *
+small_region_for_ptr_no_lock(szone_t *szone, const void *ptr)
+{
+  return hash_lookup_region_no_lock(szone->small_regions,
+                                    szone->num_small_regions_allocated,
+                                    SMALL_REGION_FOR_PTR(ptr));

 }
 
 static INLINE void
 }
 
 static INLINE void

-small_free_no_lock(szone_t *szone, small_region_t *region, void *ptr, msize_t msize) {
+small_free_no_lock(szone_t *szone, region_t *region, void *ptr, msize_t msize)
+{
+    msize_t            *meta_headers = SMALL_META_HEADER_FOR_PTR(ptr);
+    unsigned           index = SMALL_META_INDEX_FOR_PTR(ptr);
+    size_t             original_size = SMALL_BYTES_FOR_MSIZE(msize);
+    unsigned char      *next_block = ((unsigned char *)ptr + original_size);
+    msize_t            next_index = index + msize;
+    msize_t            previous_msize, next_msize;
+    void               *previous;
+

     // Assumes locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     // Assumes locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);

-    msize_t    *meta_headers = small_meta_headers(ptr);
-    msize_t    index = small_meta_index(ptr);
-    size_t     original_size = msize << SHIFT_SMALL_QUANTUM;
-    void       *next_block = ((char *)ptr + original_size);
-    msize_t    next_index = index + msize;

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In small_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);
+       malloc_printf("in small_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);

     }
     if (! msize) {
     }
     if (! msize) {

-       malloc_printf("In small_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);
-       szone_error(szone, "Trying to free small block that is too small", ptr);
+       szone_error(szone, "trying to free small block that is too small", ptr,
+           "in small_free_no_lock(), ptr=%p, msize=%d\n", ptr, msize);

     }
     }

-    // printf("In small_free_no_lock %p - msize=%d\n", ptr, msize);

 #endif
     // We try to coalesce this block with the preceeding one
     if (index && (SMALL_PREVIOUS_MSIZE(ptr) <= index)) {
 #endif
     // We try to coalesce this block with the preceeding one
     if (index && (SMALL_PREVIOUS_MSIZE(ptr) <= index)) {

-       msize_t previous_msize = SMALL_PREVIOUS_MSIZE(ptr);
+       previous_msize = SMALL_PREVIOUS_MSIZE(ptr);

        if (meta_headers[index - previous_msize] == (previous_msize | SMALL_IS_FREE)) {
        if (meta_headers[index - previous_msize] == (previous_msize | SMALL_IS_FREE)) {

-           void        *previous = ptr - (previous_msize << SHIFT_SMALL_QUANTUM);
+           previous = ptr - SMALL_BYTES_FOR_MSIZE(previous_msize);

            // previous is really to be coalesced
 #if DEBUG_MALLOC
            if (LOG(szone, ptr) || LOG(szone,previous)) { 
            // previous is really to be coalesced
 #if DEBUG_MALLOC
            if (LOG(szone, ptr) || LOG(szone,previous)) { 

-               malloc_printf("In small_free_no_lock(), coalesced backwards for %p previous=%p\n", ptr, previous);
+               malloc_printf("in small_free_no_lock(), coalesced backwards for %p previous=%p\n", ptr, previous);

            }
 #endif
            }
 #endif

-           // malloc_printf("In small_free_no_lock(), coalesced backwards for %p previous=%p\n", ptr, previous);

            small_free_list_remove_ptr(szone, previous, previous_msize);
            small_meta_header_set_middle(meta_headers, index);
            ptr = previous;
            small_free_list_remove_ptr(szone, previous, previous_msize);
            small_meta_header_set_middle(meta_headers, index);
            ptr = previous;


@@ -1498,22 +2112,21 @@ small_free_no_lock(szone_t *szone, small_region_t *region, void *ptr, msize_t ms
        }
     }
     // We try to coalesce with the next block
        }
     }
     // We try to coalesce with the next block

-    if (((vm_address_t)next_block < SMALL_REGION_END(*region)) && (meta_headers[next_index] & SMALL_IS_FREE)) {
+    if ((next_block < SMALL_REGION_END(*region)) && (meta_headers[next_index] & SMALL_IS_FREE)) {

        // next block is free, we coalesce
        // next block is free, we coalesce

-       msize_t next_msize = meta_headers[next_index] & ~ SMALL_IS_FREE;
+       next_msize = meta_headers[next_index] & ~ SMALL_IS_FREE;

 #if DEBUG_MALLOC
        if (LOG(szone,ptr)) malloc_printf("In small_free_no_lock(), for ptr=%p, msize=%d coalesced next block=%p next_msize=%d\n", ptr, msize, next_block, next_msize);
 #endif
 #if DEBUG_MALLOC
        if (LOG(szone,ptr)) malloc_printf("In small_free_no_lock(), for ptr=%p, msize=%d coalesced next block=%p next_msize=%d\n", ptr, msize, next_block, next_msize);
 #endif

-       // malloc_printf("In small_free_no_lock(), for ptr=%p, msize=%d coalesced next block=%p next_msize=%d\n", ptr, msize, next_block, next_msize);

        small_free_list_remove_ptr(szone, next_block, next_msize);
        small_meta_header_set_middle(meta_headers, next_index);
        msize += next_msize;
     }
     if (szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) {
        if (!msize) {
        small_free_list_remove_ptr(szone, next_block, next_msize);
        small_meta_header_set_middle(meta_headers, next_index);
        msize += next_msize;
     }
     if (szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) {
        if (!msize) {

-           szone_error(szone, "Incorrect size information - block header was damaged", ptr);
+           szone_error(szone, "incorrect size information - block header was damaged", ptr, NULL);

        } else {
        } else {

-           memset(ptr, 0x55, (msize << SHIFT_SMALL_QUANTUM));
+           memset(ptr, 0x55, SMALL_BYTES_FOR_MSIZE(msize));

        }
     }
     small_free_list_add_ptr(szone, ptr, msize);
        }
     }
     small_free_list_add_ptr(szone, ptr, msize);


@@ -1523,110 +2136,206 @@ small_free_no_lock(szone_t *szone, small_region_t *region, void *ptr, msize_t ms
 }
 
 static void *
 }
 
 static void *

-small_malloc_from_region_no_lock(szone_t *szone, msize_t msize) {
+small_malloc_from_region_no_lock(szone_t *szone, msize_t msize)
+{
+    void               *last_block;
+    void               *ptr;
+    void               *new_address;
+    msize_t            *meta_headers;
+    msize_t            index ;
+    msize_t            msize_left;
+

     // Allocates from the last region or a freshly allocated region
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     // Before anything we transform the small_bytes_free_at_end - if any - to a regular free block
     if (szone->small_bytes_free_at_end) {
     // Allocates from the last region or a freshly allocated region
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     // Before anything we transform the small_bytes_free_at_end - if any - to a regular free block
     if (szone->small_bytes_free_at_end) {

-       small_region_t  last_region = szone->small_regions[szone->num_small_regions - 1];
-       void    *last_block = (void *)(SMALL_REGION_END(last_region) - szone->small_bytes_free_at_end);
-       small_free_list_add_ptr(szone, last_block, szone->small_bytes_free_at_end >> SHIFT_SMALL_QUANTUM);
-       small_meta_header(last_block)[0] = (szone->small_bytes_free_at_end >> SHIFT_SMALL_QUANTUM) | SMALL_IS_FREE;
-       szone->small_bytes_free_at_end = 0;
+      last_block = (void *)(SMALL_REGION_END(szone->last_small_region) - szone->small_bytes_free_at_end);
+      small_free_list_add_ptr(szone, last_block, SMALL_MSIZE_FOR_BYTES(szone->small_bytes_free_at_end));
+      *SMALL_METADATA_FOR_PTR(last_block) = SMALL_MSIZE_FOR_BYTES(szone->small_bytes_free_at_end) | SMALL_IS_FREE;
+      szone->small_bytes_free_at_end = 0;

     }
     }

-    void               *ptr;

     // time to create a new region
     // time to create a new region

-    vm_address_t       new_address = allocate_pages(szone, SMALL_REGION_SIZE, SMALL_BLOCKS_ALIGN, 0, VM_MAKE_TAG(VM_MEMORY_MALLOC_SMALL));
-    if (!new_address) {
-       // out of memory!
-       return NULL;
-    }
-    ptr = (void *)new_address;
-    msize_t    *meta_headers = small_meta_headers(ptr);
-    msize_t    index = 0;
-    // malloc_printf("Allocated small region #%d: %p [%y]\n", szone->num_small_regions, new_address, SMALL_REGION_SIZE);
-    if (szone->num_small_regions == INITIAL_NUM_SMALL_REGIONS) {
-       // time to grow the number of regions
-       unsigned        region_capacity = (1 << (32 - SMALL_BLOCKS_ALIGN)) - 20; // that is for sure the maximum number of small regions we can have
-       msize_t         new_msize = (region_capacity * sizeof(small_region_t) + SMALL_QUANTUM - 1) / SMALL_QUANTUM;
-       small_region_t  *new_regions = ptr;
-       // malloc_printf("Now %d small_regions growing regions %p to %d msize=%d\n", szone->num_small_regions + 1, szone->small_regions, region_capacity, new_msize);
-       small_meta_header_set_in_use(meta_headers, index, new_msize);
-       szone->num_small_objects++;
-       szone->num_bytes_in_small_objects += new_msize << SHIFT_SMALL_QUANTUM;
-       memcpy(new_regions, szone->small_regions, INITIAL_NUM_SMALL_REGIONS * sizeof(small_region_t));
-       // We intentionally leak the previous regions pointer to avoid multi-threading crashes if another thread was reading it (unlocked) while we are changing it.  
-       szone->small_regions = new_regions; // note we set this pointer after it's all set
-       ptr += new_msize << SHIFT_SMALL_QUANTUM;
-       index = new_msize;
-       // malloc_printf("Regions is now %p next ptr is %p\n", szone->small_regions, ptr);
-    }
-    szone->small_regions[szone->num_small_regions] = new_address >> SMALL_BLOCKS_ALIGN;
-    szone->num_small_regions++; // we bump the number of regions AFTER we have changes the regions pointer to enable finding a small region without taking the lock
-    // malloc_printf("Now %d small regions\n", szone->num_small_regions);
+    new_address = allocate_pages(szone, SMALL_REGION_SIZE, SMALL_BLOCKS_ALIGN,
+                                 0, VM_MEMORY_MALLOC_SMALL);
+    if (!new_address)
+      return NULL;
+
+    ptr = new_address;
+    meta_headers = SMALL_META_HEADER_FOR_PTR(ptr);
+    index = 0;
+    
+    // Check to see if the hash ring of small regions needs to grow.  Try to
+    // avoid the hash ring becoming too dense.
+    if (szone->num_small_regions_allocated < (2 * szone->num_small_regions)) {
+      region_t *new_regions;
+      size_t new_size;
+      new_regions = hash_regions_grow_no_lock(szone, szone->small_regions,
+                                              szone->num_small_regions_allocated,
+                                              &new_size);
+      // Do not deallocate the current small_regions allocation since someone
+         // may be iterating it.  Instead, just leak it.
+      szone->small_regions = new_regions;
+      szone->num_small_regions_allocated = new_size;
+    }
+    // Insert the new region into the hash ring, and update malloc statistics
+    hash_region_insert_no_lock(szone->small_regions, 
+                               szone->num_small_regions_allocated,
+                               new_address);
+    szone->last_small_region = new_address;
+                               
+    // we bump the number of regions AFTER we have changes the regions pointer 
+    // to enable finding a small region without taking the lock
+    //
+    // FIXME: naive assumption assumes memory ordering coherence between this
+    // and other CPUs.  This also applies to the near-identical code in
+    // tiny_malloc_from_region_no_lock.
+    szone->num_small_regions++;

     small_meta_header_set_in_use(meta_headers, index, msize);
     small_meta_header_set_in_use(meta_headers, index, msize);

-    msize_t            msize_left = NUM_SMALL_BLOCKS - index;
+    msize_left = NUM_SMALL_BLOCKS - index;

     szone->num_small_objects++;
     szone->num_small_objects++;

-    szone->num_bytes_in_small_objects += msize << SHIFT_SMALL_QUANTUM;
+    szone->num_bytes_in_small_objects += SMALL_BYTES_FOR_MSIZE(msize);

     // add a big free block
     index += msize; msize_left -= msize;
     meta_headers[index] = msize_left;
     // add a big free block
     index += msize; msize_left -= msize;
     meta_headers[index] = msize_left;

-    szone->small_bytes_free_at_end = msize_left << SHIFT_SMALL_QUANTUM;
-    // malloc_printf("small_bytes_free_at_end set to %d\n", szone-> small_bytes_free_at_end);
+    szone->small_bytes_free_at_end = SMALL_BYTES_FOR_MSIZE(msize_left);

     return ptr;
 }
 
     return ptr;
 }
 

+static INLINE boolean_t
+try_realloc_small_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size)
+{
+    // returns 1 on success
+    msize_t    *meta_headers = SMALL_META_HEADER_FOR_PTR(ptr);
+    unsigned   index = SMALL_META_INDEX_FOR_PTR(ptr);
+    msize_t    old_msize = SMALL_MSIZE_FOR_BYTES(old_size);
+    msize_t    new_msize = SMALL_MSIZE_FOR_BYTES(new_size + SMALL_QUANTUM - 1);
+    void       *next_block = (char *)ptr + old_size;
+    unsigned   next_index = index + old_msize;
+    msize_t    next_msize_and_free;
+    msize_t    next_msize;
+    msize_t    leftover_msize;
+    void       *leftover;
+    unsigned   leftover_index;
+
+    if (next_index >= NUM_SMALL_BLOCKS) {
+       return 0;
+    }
+#if DEBUG_MALLOC
+    if ((uintptr_t)next_block & (SMALL_QUANTUM - 1)) {
+       szone_error(szone, "internal invariant broken in realloc(next_block)", next_block, NULL);
+    }
+    if (meta_headers[index] != old_msize)
+       malloc_printf("*** try_realloc_small_in_place incorrect old %d %d\n",
+         meta_headers[index], old_msize);
+#endif
+    SZONE_LOCK(szone);
+    /*
+     * Look for a free block immediately afterwards.  If it's large enough, we can consume (part of)
+     * it.
+     */
+    next_msize_and_free = meta_headers[next_index];
+    next_msize = next_msize_and_free & ~ SMALL_IS_FREE;
+    if (!(next_msize_and_free & SMALL_IS_FREE) || (old_msize + next_msize < new_msize)) {
+       SZONE_UNLOCK(szone);
+       return 0;
+    }
+    /*
+     * The following block is big enough; pull it from its freelist and chop off enough to satisfy
+     * our needs.
+     */
+    small_free_list_remove_ptr(szone, next_block, next_msize);
+    small_meta_header_set_middle(meta_headers, next_index);
+    leftover_msize = old_msize + next_msize - new_msize;
+    if (leftover_msize) {
+       /* there's some left, so put the remainder back */
+       leftover = (unsigned char *)ptr + SMALL_BYTES_FOR_MSIZE(new_msize);
+       small_free_list_add_ptr(szone, leftover, leftover_msize);
+       leftover_index = index + new_msize;
+       small_meta_header_set_is_free(meta_headers, leftover_index, leftover_msize);
+    }
+#if DEBUG_MALLOC
+    if (SMALL_BYTES_FOR_MSIZE(new_msize) >= LARGE_THRESHOLD) {
+       malloc_printf("*** realloc in place for %p exceeded msize=%d\n", new_msize);
+    }
+#endif
+    small_meta_header_set_in_use(meta_headers, index, new_msize);
+#if DEBUG_MALLOC
+    if (LOG(szone,ptr)) {
+       malloc_printf("in szone_realloc(), ptr=%p, msize=%d\n", ptr, *SMALL_METADATA_FOR_PTR(ptr));
+    }
+#endif
+    szone->num_bytes_in_small_objects += SMALL_BYTES_FOR_MSIZE(new_msize - old_msize);
+    SZONE_UNLOCK(szone);
+    CHECK(szone, __PRETTY_FUNCTION__);
+    return 1;
+}
+

 static boolean_t
 static boolean_t

-szone_check_small_region(szone_t *szone, small_region_t *region) {
-    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
-    void               *ptr = (void *)SMALL_REGION_ADDRESS(*region);
-    msize_t            *meta_headers = small_meta_headers(ptr);
-    vm_address_t       region_end = SMALL_REGION_END(*region);
+szone_check_small_region(szone_t *szone, region_t region)
+{
+    unsigned char      *ptr = SMALL_REGION_ADDRESS(region);
+    msize_t            *meta_headers = SMALL_META_HEADER_FOR_PTR(ptr);
+    unsigned char      *region_end = SMALL_REGION_END(region);

     msize_t            prev_free = 0;
     msize_t            prev_free = 0;

-    if (region == szone->small_regions + szone->num_small_regions - 1) region_end -= szone->small_bytes_free_at_end;
-    while ((vm_address_t)ptr < region_end) {
-       msize_t         index = small_meta_index(ptr);
-       msize_t         msize_and_free = meta_headers[index];
-       if (! (msize_and_free & SMALL_IS_FREE)) {
+    unsigned           index;
+    msize_t            msize_and_free;
+    msize_t            msize;
+    free_list_t                *free_head;
+    void            *previous, *next;
+    msize_t            *follower;
+
+    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
+    if (region == szone->last_small_region) region_end -= szone->small_bytes_free_at_end;
+    while (ptr < region_end) {
+       index = SMALL_META_INDEX_FOR_PTR(ptr);
+       msize_and_free = meta_headers[index];
+       if (!(msize_and_free & SMALL_IS_FREE)) {

            // block is in use
            // block is in use

-           msize_t     msize = msize_and_free;
+           msize = msize_and_free;

            if (!msize) {
            if (!msize) {

-               malloc_printf("*** malloc[%d]: invariant broken: null msize ptr=%p region#=%d num_small_regions=%d end=%p\n", getpid(), ptr, region - szone->small_regions, szone->num_small_regions, (void *)region_end);
+               malloc_printf("*** invariant broken: null msize ptr=%p num_small_regions=%d end=%p\n",
+                 ptr, szone->num_small_regions, region_end);

                return 0;
            }
            if (msize > (LARGE_THRESHOLD / SMALL_QUANTUM)) {
                return 0;
            }
            if (msize > (LARGE_THRESHOLD / SMALL_QUANTUM)) {

-               malloc_printf("*** malloc[%d]: invariant broken for %p this small msize=%d - size is too large\n", getpid(), ptr, msize_and_free);
+               malloc_printf("*** invariant broken for %p this small msize=%d - size is too large\n",
+                 ptr, msize_and_free);

                return 0;
            }
                return 0;
            }

-           ptr += msize << SHIFT_SMALL_QUANTUM;
+           ptr += SMALL_BYTES_FOR_MSIZE(msize);

            prev_free = 0;
        } else {
            // free pointer
            prev_free = 0;
        } else {
            // free pointer

-           msize_t     msize = msize_and_free & ~ SMALL_IS_FREE;
-           free_list_t *free_head = ptr;
-           msize_t     *follower = (void *)FOLLOWING_SMALL_PTR(ptr, msize);
-           if (! msize) {
-               malloc_printf("*** malloc[%d]: invariant broken for free block %p this msize=%d\n", getpid(), ptr, msize);
+           msize = msize_and_free & ~ SMALL_IS_FREE;
+           free_head = (free_list_t *)ptr;
+           follower = (msize_t *)FOLLOWING_SMALL_PTR(ptr, msize);
+           if (!msize) {
+               malloc_printf("*** invariant broken for free block %p this msize=%d\n", ptr, msize);

                return 0;
            }
            if (prev_free) {
                return 0;
            }
            if (prev_free) {

-               malloc_printf("*** malloc[%d]: invariant broken for %p (2 free in a row)\n", getpid(), ptr);
+               malloc_printf("*** invariant broken for %p (2 free in a row)\n", ptr);

                return 0;
            }
            free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);
                return 0;
            }
            free_list_checksum(szone, free_head, __PRETTY_FUNCTION__);

-           if (free_head->previous && !(small_meta_header(free_head->previous)[0] & SMALL_IS_FREE)) {
-               malloc_printf("*** malloc[%d]: invariant broken for %p (previous %p is not a free pointer)\n", getpid(), ptr, free_head->previous);
+           previous = free_list_unchecksum_ptr(free_head->previous);
+           next = free_list_unchecksum_ptr(free_head->next);
+           if (previous && !SMALL_PTR_IS_FREE(previous)) {
+               malloc_printf("*** invariant broken for %p (previous %p is not a free pointer)\n",
+                 ptr, free_head->previous);

                return 0;
            }
                return 0;
            }

-           if (free_head->next && !(small_meta_header(free_head->next)[0] & SMALL_IS_FREE)) {
-               malloc_printf("*** malloc[%d]: invariant broken for %p (next is not a free pointer)\n", getpid(), ptr);
+           if (next && !SMALL_PTR_IS_FREE(next)) {
+               malloc_printf("*** invariant broken for %p (next is not a free pointer)\n", ptr);

                return 0;
            }
            if (SMALL_PREVIOUS_MSIZE(follower) != msize) {
                return 0;
            }
            if (SMALL_PREVIOUS_MSIZE(follower) != msize) {

-               malloc_printf("*** malloc[%d]: invariant broken for small free %p followed by %p in region [%p-%p] (end marker incorrect) should be %d; in fact %d\n", getpid(), ptr, follower, SMALL_REGION_ADDRESS(*region), region_end, msize, SMALL_PREVIOUS_MSIZE(follower));
+               malloc_printf("*** invariant broken for small free %p followed by %p in region [%p-%p] "
+                 "(end marker incorrect) should be %d; in fact %d\n",
+                 ptr, follower, SMALL_REGION_ADDRESS(region), region_end, msize, SMALL_PREVIOUS_MSIZE(follower));

                return 0;
            }
                return 0;
            }

-           ptr = follower;
+           ptr = (unsigned char *)follower;

            prev_free = SMALL_IS_FREE;
        }
     }
            prev_free = SMALL_IS_FREE;
        }
     }


@@ -1634,174 +2343,229 @@ szone_check_small_region(szone_t *szone, small_region_t *region) {
 }
 
 static kern_return_t
 }
 
 static kern_return_t

-small_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t region_address, unsigned short num_regions, size_t small_bytes_free_at_end, memory_reader_t reader, vm_range_recorder_t recorder) {
-    small_region_t     *regions;
-    unsigned           index = 0;
-    vm_range_t         buffer[MAX_RECORDER_BUFFER];
-    unsigned           count = 0;
-    kern_return_t      err;
-    err = reader(task, region_address, sizeof(small_region_t) * num_regions, (void **)&regions);
-    if (err) return err;
-    while (index < num_regions) {
-       small_region_t  region = regions[index];
-       vm_range_t      range = {SMALL_REGION_ADDRESS(region), SMALL_REGION_SIZE};
-       // malloc_printf("Enumerating small ptrs for Region starting at %p\n", range.address);
-       if (type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) {
-           vm_range_t  admin_range = {range.address + (1 << SMALL_BLOCKS_ALIGN), range.size - (1 << SMALL_BLOCKS_ALIGN)};
-           recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &admin_range, 1);
-       }
-       if (type_mask & (MALLOC_PTR_REGION_RANGE_TYPE | MALLOC_ADMIN_REGION_RANGE_TYPE)) {
-           vm_range_t  ptr_range = {range.address, 1 << SMALL_BLOCKS_ALIGN};
-           recorder(task, context, MALLOC_PTR_REGION_RANGE_TYPE, &ptr_range, 1);
-       }
-       if (type_mask & MALLOC_PTR_IN_USE_RANGE_TYPE) {
-           unsigned char       *mapped_region;
-           err = reader(task, range.address, range.size, (void **)&mapped_region);
-           if (err) return err;
-           msize_t     *block_header = (msize_t *)(mapped_region + (1 << SMALL_BLOCKS_ALIGN));
-           unsigned            block_index = 0;
-           unsigned            block_limit = NUM_SMALL_BLOCKS;
-           if (index == num_regions - 1)
-               block_limit -= (small_bytes_free_at_end >> SHIFT_SMALL_QUANTUM);
-           while (block_index < block_limit) {
-               msize_t msize_and_free = block_header[block_index];
-               msize_t msize = msize_and_free & ~ SMALL_IS_FREE;
-               if (! (msize_and_free & SMALL_IS_FREE)) {
-                   // Block in use
-                   buffer[count].address = range.address + (block_index << SHIFT_SMALL_QUANTUM);
-                   buffer[count].size = msize << SHIFT_SMALL_QUANTUM;
-                   count++;
-                   if (count >= MAX_RECORDER_BUFFER) {
-                       recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
-                       count = 0;
-                   }
-               }
-               block_index += msize;
-           }
-           // malloc_printf("End small region - count=%d\n", count);
-       }
-       index++;
-    }
-    if (count) {
-       recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
-    }
-    return 0;
+small_in_use_enumerator(task_t task, void *context, unsigned type_mask, szone_t *szone, memory_reader_t reader, vm_range_recorder_t recorder)
+{
+  size_t num_regions = szone->num_small_regions_allocated;
+  void *last_small_free = szone->last_small_free; 
+  size_t       index;
+  region_t     *regions;
+  vm_range_t           buffer[MAX_RECORDER_BUFFER];
+  unsigned             count = 0;
+  kern_return_t        err;
+  region_t     region;
+  vm_range_t           range;
+  vm_range_t           admin_range;
+  vm_range_t           ptr_range;
+  unsigned char        *mapped_region;
+  msize_t              *block_header;
+  unsigned             block_index;
+  unsigned             block_limit;
+  msize_t              msize_and_free;
+  msize_t              msize;
+  vm_address_t last_small_free_ptr = 0;
+  msize_t last_small_free_msize = 0;
+  
+  if (last_small_free) {
+    last_small_free_ptr = (uintptr_t)last_small_free & ~(SMALL_QUANTUM - 1);
+    last_small_free_msize = (uintptr_t)last_small_free & (SMALL_QUANTUM - 1);
+  }
+  
+  err = reader(task, (vm_address_t)szone->small_regions, sizeof(region_t) * num_regions, (void **)&regions);
+  if (err) return err;
+  for (index = 0; index < num_regions; ++index) {
+    region = regions[index];
+    if (region) {
+      range.address = (vm_address_t)SMALL_REGION_ADDRESS(region);
+      range.size = SMALL_REGION_SIZE;
+      if (type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) {
+        admin_range.address = range.address + SMALL_HEADER_START;
+        admin_range.size = SMALL_ARRAY_SIZE;
+        recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &admin_range, 1);
+      }
+      if (type_mask & (MALLOC_PTR_REGION_RANGE_TYPE | MALLOC_ADMIN_REGION_RANGE_TYPE)) {
+        ptr_range.address = range.address;
+        ptr_range.size = NUM_SMALL_BLOCKS * SMALL_QUANTUM;
+        recorder(task, context, MALLOC_PTR_REGION_RANGE_TYPE, &ptr_range, 1);
+      }
+      if (type_mask & MALLOC_PTR_IN_USE_RANGE_TYPE) {
+        err = reader(task, range.address, range.size, (void **)&mapped_region);
+        if (err) return err;
+        block_header = (msize_t *)(mapped_region + SMALL_HEADER_START);
+        block_index = 0;
+        block_limit = NUM_SMALL_BLOCKS;
+        if (region == szone->last_small_region)
+          block_limit -= SMALL_MSIZE_FOR_BYTES(szone->small_bytes_free_at_end);
+        while (block_index < block_limit) {
+          msize_and_free = block_header[block_index];
+          msize = msize_and_free & ~ SMALL_IS_FREE;
+          if (! (msize_and_free & SMALL_IS_FREE) &&
+              range.address + SMALL_BYTES_FOR_MSIZE(block_index) != last_small_free_ptr) {
+            // Block in use
+            buffer[count].address = range.address + SMALL_BYTES_FOR_MSIZE(block_index);
+            buffer[count].size = SMALL_BYTES_FOR_MSIZE(msize);
+            count++;
+            if (count >= MAX_RECORDER_BUFFER) {
+              recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
+              count = 0;
+            }
+          }
+          block_index += msize;
+        }
+      }
+    }
+  }
+  if (count) {
+    recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE, buffer, count);
+  }
+  return 0;

 }
 
 }
 

-static INLINE void *
-small_malloc_from_free_list(szone_t *szone, msize_t msize) {
+static void *
+small_malloc_from_free_list(szone_t *szone, msize_t msize)
+{
+    free_list_t                *ptr;
+    msize_t         this_msize;
+    grain_t         slot = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;
+    free_list_t                **free_list = szone->small_free_list;
+    free_list_t     *next;
+    free_list_t     **limit;
+    unsigned        bitmap = szone->small_bitmap & ~ ((1 << slot) - 1);
+    msize_t         leftover_msize;
+    free_list_t     *leftover_ptr;
+    msize_t         *meta_headers;
+    unsigned        leftover_index;
+

     // Assumes locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     // Assumes locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);

-    grain_t    grain = (msize <= NUM_SMALL_SLOTS) ? msize - 1 : NUM_SMALL_SLOTS - 1;
-    unsigned   bitmap = szone->small_bitmap & ~ ((1 << grain) - 1);
-    void       *ptr;
-    msize_t    this_msize;
-    if (!bitmap) goto try_small_from_end;
-    grain = BITMAP32_FFS(bitmap) - 1;
-    // first try the small grains
-    free_list_t        **free_list;
-    free_list_t        **limit = szone->small_free_list + NUM_SMALL_SLOTS - 1;
-    free_list = szone->small_free_list + grain;
+    
+    // Mask off the bits representing slots holding free blocks smaller than the
+    // size we need.  If there are no larger free blocks, try allocating from
+    // the free space at the end of the tiny region.
+    if (!bitmap) 
+        goto try_small_from_end;
+        
+    slot = BITMAP32_CTZ(bitmap);
+    limit = free_list + NUM_SMALL_SLOTS - 1;
+    free_list += slot;
+
+    // Iterate over freelists looking for free blocks, starting at first list
+    // which is not empty, and contains blocks which are large enough to satisfy
+    // our request.

     while (free_list < limit) {
     while (free_list < limit) {

-       // try bigger grains
-       ptr = *free_list;
-       if (ptr) {
-           free_list_t *next;
-           next = ((free_list_t *)ptr)->next;
-           if (next) {
-               next->previous = NULL;
-               free_list_set_checksum(szone, next);
-           }
-           *free_list = next;
-           this_msize = small_meta_header(ptr)[0] & ~ SMALL_IS_FREE;
-           // malloc_printf("small_malloc_from_free_list: allocated from free list\n");
-           goto add_leftover_and_proceed;
-       }
-       free_list++;
-    }
-    // We now check the large grains for one that is big enough
-    ptr = *free_list;
+        ptr = *free_list;
+        if (ptr) {
+            next = free_list_unchecksum_ptr(ptr->next);
+            *free_list = next;
+            this_msize = SMALL_PTR_SIZE(ptr);
+            if (next) {
+                next->previous = ptr->previous;
+            } else {
+                BITMAP32_CLR(szone->small_bitmap, this_msize - 1);
+            }
+            goto add_leftover_and_proceed;
+        }
+        free_list++;
+    }
+
+    // We are now looking at the last slot, which contains blocks equal to, or
+    // due to coalescing of free blocks, larger than 31 * small quantum size.
+    // If the last freelist is not empty, and the head contains a block that is
+    // larger than our request, then the remainder is put back on the free list.
+    //
+    // FIXME: This code doesn't have the optimization from the 'tiny' codepath 
+    //        that optimizes for the this_msize >= 2 * num slots
+    // FIXME: this code also seems somewhat bogus.  There's a check for
+    //        this_msize >= msize, but by definition we can't ask for a small
+    //        block larger than 31 small quanta, and every free block in this
+    //        slot has to be at least that large.
+    ptr = *limit;

     while (ptr) {
     while (ptr) {

-       this_msize = small_meta_header(ptr)[0] & ~ SMALL_IS_FREE;
-       if (this_msize >= msize) {
-           // malloc_printf("small_malloc_from_free_list: allocated from last free list\n");
-           small_free_list_remove_ptr(szone, ptr, this_msize);
-           goto add_leftover_and_proceed;
-       }
-       ptr = ((free_list_t *)ptr)->next;
+        free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
+        next = free_list_unchecksum_ptr(ptr->next);
+        this_msize = SMALL_PTR_SIZE(ptr);
+        if (this_msize >= msize) {
+            small_free_list_remove_ptr(szone, ptr, this_msize);
+            goto add_leftover_and_proceed;
+        }
+        ptr = next;

     }
     }

+    

 try_small_from_end:
     // Let's see if we can use szone->small_bytes_free_at_end
 try_small_from_end:
     // Let's see if we can use szone->small_bytes_free_at_end

-    // malloc_printf("Found nothing in free list small_bytes_free_at_end=%y\n", szone-> small_bytes_free_at_end);
-    if (szone->small_bytes_free_at_end >= (msize << SHIFT_SMALL_QUANTUM)) {
-       ptr = (void *)(SMALL_REGION_END(szone->small_regions[szone->num_small_regions-1]) - szone->small_bytes_free_at_end);
-       szone->small_bytes_free_at_end -= msize << SHIFT_SMALL_QUANTUM;
-       if (szone->small_bytes_free_at_end) {
-           // let's mark this block as in use to serve as boundary
-           small_meta_header(ptr + (msize << SHIFT_SMALL_QUANTUM))[0] = szone->small_bytes_free_at_end >> SHIFT_SMALL_QUANTUM;
-       }
-       this_msize = msize;
-       goto return_small_alloc;
+    if (szone->small_bytes_free_at_end >= SMALL_BYTES_FOR_MSIZE(msize)) {
+        ptr = (free_list_t *)(SMALL_REGION_END(szone->last_small_region) - szone->small_bytes_free_at_end);
+        szone->small_bytes_free_at_end -= SMALL_BYTES_FOR_MSIZE(msize);
+        if (szone->small_bytes_free_at_end) {
+            // let's mark this block as in use to serve as boundary
+            *SMALL_METADATA_FOR_PTR((unsigned char *)ptr + SMALL_BYTES_FOR_MSIZE(msize)) = SMALL_MSIZE_FOR_BYTES(szone->small_bytes_free_at_end);
+        }
+        this_msize = msize;
+        goto return_small_alloc;

     }
     return NULL;
     }
     return NULL;

+    

 add_leftover_and_proceed:
     if (this_msize > msize) {
 add_leftover_and_proceed:
     if (this_msize > msize) {

-       msize_t leftover_msize = this_msize - msize;
-       void    *leftover_ptr = ptr + (msize << SHIFT_SMALL_QUANTUM);
+        leftover_msize = this_msize - msize;
+        leftover_ptr = (free_list_t *)((unsigned char *)ptr + SMALL_BYTES_FOR_MSIZE(msize));

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone,ptr)) {
-           malloc_printf("In small_malloc_from_free_list(), adding leftover ptr=%p, this_msize=%d\n", ptr, this_msize);
-       }
+        if (LOG(szone,ptr)) {
+            malloc_printf("in small_malloc_from_free_list(), adding leftover ptr=%p, this_msize=%d\n", ptr, this_msize);
+        }

 #endif
 #endif

-       small_free_list_add_ptr(szone, leftover_ptr, leftover_msize);
-       msize_t *meta_headers = small_meta_headers(leftover_ptr);
-       msize_t leftover_index = small_meta_index(leftover_ptr);
-       small_meta_header_set_is_free(meta_headers, leftover_index, leftover_msize);
-       this_msize = msize;
+        small_free_list_add_ptr(szone, leftover_ptr, leftover_msize);
+        meta_headers = SMALL_META_HEADER_FOR_PTR(leftover_ptr);
+        leftover_index = SMALL_META_INDEX_FOR_PTR(leftover_ptr);
+        small_meta_header_set_is_free(meta_headers, leftover_index, leftover_msize);
+        this_msize = msize;

     }
     }

+    

 return_small_alloc:
     szone->num_small_objects++;
 return_small_alloc:
     szone->num_small_objects++;

-    szone->num_bytes_in_small_objects += this_msize << SHIFT_SMALL_QUANTUM;
+    szone->num_bytes_in_small_objects += SMALL_BYTES_FOR_MSIZE(this_msize);

 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone,ptr)) {

-       malloc_printf("In small_malloc_from_free_list(), ptr=%p, this_msize=%d, msize=%d\n", ptr, this_msize, msize);
+       malloc_printf("in small_malloc_from_free_list(), ptr=%p, this_msize=%d, msize=%d\n", ptr, this_msize, msize);

     }
 #endif
     }
 #endif

-    small_meta_header(ptr)[0] = this_msize;
+    *SMALL_METADATA_FOR_PTR(ptr) = this_msize;

     return ptr;
 }
 
 static INLINE void *
     return ptr;
 }
 
 static INLINE void *

-small_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested) {
+small_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_requested)
+{

     boolean_t  locked = 0;
     boolean_t  locked = 0;

+#if SMALL_CACHE

     void       *ptr;
     void       *ptr;

+#endif
+    

 #if SMALL_CACHE
     ptr = (void *)szone->last_small_free;
 #if SMALL_CACHE
     ptr = (void *)szone->last_small_free;

-    if ((((unsigned)ptr) & (SMALL_QUANTUM - 1)) == msize) {
+    if ((((uintptr_t)ptr) & (SMALL_QUANTUM - 1)) == msize) {

        // we have a candidate - let's lock to make sure
        LOCK_AND_NOTE_LOCKED(szone, locked);
        if (ptr == (void *)szone->last_small_free) {
            szone->last_small_free = NULL;
        // we have a candidate - let's lock to make sure
        LOCK_AND_NOTE_LOCKED(szone, locked);
        if (ptr == (void *)szone->last_small_free) {
            szone->last_small_free = NULL;

-           // malloc_printf("using last_small_free\n");

            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);
            SZONE_UNLOCK(szone);
            CHECK(szone, __PRETTY_FUNCTION__);

-           ptr = (void *)((unsigned)ptr & ~ (SMALL_QUANTUM - 1));
+           ptr = (void *)((uintptr_t)ptr & ~ (SMALL_QUANTUM - 1));

            if (cleared_requested) {
            if (cleared_requested) {

-               memset(ptr, 0, msize << SHIFT_SMALL_QUANTUM);
+               memset(ptr, 0, SMALL_BYTES_FOR_MSIZE(msize));

            }
            return ptr;
        }
            }
            return ptr;
        }

-       // malloc_printf("optimistic locking for last_small_free failed\n");

     }
 #endif
     }
 #endif

-    // Except in rare occasions where we need to add a new region, we are going to end up locking, so we might as well lock right away to avoid doing unnecessary optimistic probes
+    // Except in rare occasions where we need to add a new region, we are going to end up locking,
+    // so we might as well lock right away to avoid doing unnecessary optimistic probes

     if (!locked) LOCK_AND_NOTE_LOCKED(szone, locked);
     ptr = small_malloc_from_free_list(szone, msize);
     if (ptr) {
        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        if (cleared_requested) {
     if (!locked) LOCK_AND_NOTE_LOCKED(szone, locked);
     ptr = small_malloc_from_free_list(szone, msize);
     if (ptr) {
        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        if (cleared_requested) {

-           memset(ptr, 0, msize << SHIFT_SMALL_QUANTUM);
+           memset(ptr, 0, SMALL_BYTES_FOR_MSIZE(msize));

        }
        return ptr;
     }
        }
        return ptr;
     }


@@ -1812,15 +2576,17 @@ small_malloc_should_clear(szone_t *szone, msize_t msize, boolean_t cleared_reque
     return ptr;
 }
 
     return ptr;
 }
 

+// tries to allocate a small, cleared block

 static INLINE void *
 static INLINE void *

-small_malloc_cleared_no_lock(szone_t *szone, msize_t msize) {
-    // tries to allocate a small, cleared block
+small_malloc_cleared_no_lock(szone_t *szone, msize_t msize)
+{
+    void       *ptr;
+

     // Assumes already locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     // Assumes already locked
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);

-    void       *ptr;

     ptr = small_malloc_from_free_list(szone, msize);
     if (ptr) {
     ptr = small_malloc_from_free_list(szone, msize);
     if (ptr) {

-       memset(ptr, 0, msize << SHIFT_SMALL_QUANTUM);
+       memset(ptr, 0, SMALL_BYTES_FOR_MSIZE(msize));

        return ptr;
     } else {
        ptr = small_malloc_from_region_no_lock(szone, msize);
        return ptr;
     } else {
        ptr = small_malloc_from_region_no_lock(szone, msize);


@@ -1830,184 +2596,325 @@ small_malloc_cleared_no_lock(szone_t *szone, msize_t msize) {
 }
 
 static INLINE void
 }
 
 static INLINE void

-free_small(szone_t *szone, void *ptr, small_region_t *small_region) {
+free_small(szone_t *szone, void *ptr, region_t *small_region)
+{
+    msize_t            msize = SMALL_PTR_SIZE(ptr);
+#if SMALL_CACHE
+    void               *ptr2;
+#endif
+    

     // ptr is known to be in small_region
     // ptr is known to be in small_region

-    msize_t            msize_and_free;
-    msize_and_free = small_meta_header(ptr)[0];
-    if (msize_and_free & SMALL_IS_FREE) {
-       szone_error(szone, "Object already freed being freed", ptr);
+    SZONE_LOCK(szone);
+#if SMALL_CACHE
+    ptr2 = szone->last_small_free;
+    /* check that we don't already have this pointer in the cache */
+    if (ptr == (void *)((uintptr_t)ptr2 & ~ (SMALL_QUANTUM - 1))) {
+       szone_error(szone, "double free", ptr, NULL);
+       return;
+    }
+#endif
+    if (SMALL_PTR_IS_FREE(ptr)) {
+       szone_error(szone, "double free", ptr, NULL);

        return;
     }
        return;
     }

-    CHECK(szone, __PRETTY_FUNCTION__);
-    // malloc_printf("%p[%x]\n", ptr, msize_and_free);
-    SZONE_LOCK(szone);

 #if SMALL_CACHE
 #if SMALL_CACHE

-    void       *ptr2 = szone->last_small_free;
-    szone->last_small_free = (void *)(((unsigned)ptr) | msize_and_free);
+    szone->last_small_free = (void *)(((uintptr_t)ptr) | msize);

     if (!ptr2) {
     if (!ptr2) {

-       // malloc_printf("stuffing last_small_free\n");

        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        return;
     }
        SZONE_UNLOCK(szone);
        CHECK(szone, __PRETTY_FUNCTION__);
        return;
     }

-    // malloc_printf("replacing previous last_small_free %p with %p\n", ptr2, szone->last_small_free);
-    msize_and_free = (unsigned)ptr2 & (SMALL_QUANTUM - 1);
-    ptr = (void *)(((unsigned)ptr2) & ~ (SMALL_QUANTUM - 1));
+    msize = (uintptr_t)ptr2 & (SMALL_QUANTUM - 1);
+    ptr = (void *)(((uintptr_t)ptr2) & ~ (SMALL_QUANTUM - 1));

     small_region = small_region_for_ptr_no_lock(szone, ptr);
     if (!small_region) {
     small_region = small_region_for_ptr_no_lock(szone, ptr);
     if (!small_region) {

-       szone_error(szone, "Double free (small cache)", ptr);
+       szone_error(szone, "double free (small cache)", ptr, NULL);
+       return;

     }
 #endif
     }
 #endif

-    small_free_no_lock(szone, small_region, ptr, msize_and_free);
+    small_free_no_lock(szone, small_region, ptr, msize);

     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
 }
 
 static void
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
 }
 
 static void

-print_small_free_list(szone_t *szone) {
+print_small_free_list(szone_t *szone)
+{

     grain_t            grain = 0;
     grain_t            grain = 0;

-    malloc_printf("Small free sizes: ");
-    while (grain < NUM_SMALL_SLOTS) {
-       free_list_t     *ptr = szone->small_free_list[grain];
-       if (ptr) {
-           malloc_printf("%s%y[%d]; ", (grain == NUM_SMALL_SLOTS-1) ? ">=" : "", (grain + 1) * SMALL_QUANTUM, free_list_count(ptr));
+    free_list_t                *ptr;
+    _SIMPLE_STRING     b = _simple_salloc();
+    
+    if (b) {
+       _simple_sappend(b, "small free sizes: ");
+       while (grain < NUM_SMALL_SLOTS) {
+           ptr = szone->small_free_list[grain];
+           if (ptr) {
+               _simple_sprintf(b, "%s%y[%d]; ", (grain == NUM_SMALL_SLOTS-1) ? ">=" : "", (grain + 1) * SMALL_QUANTUM, free_list_count(ptr));
+           }
+           grain++;

        }
        }

-       grain++;
+       _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX, "%s\n", _simple_string(b));
+       _simple_sfree(b);

     }
     }

-    malloc_printf("\n");

 }
 
 static void
 }
 
 static void

-print_small_region(szone_t *szone, boolean_t verbose, small_region_t *region, size_t bytes_at_end) {
-    unsigned   counts[1024];
-    unsigned   in_use = 0;
-    vm_address_t       start = SMALL_REGION_ADDRESS(*region);
-    vm_address_t       limit = SMALL_REGION_END(*region) - bytes_at_end;
+print_small_region(szone_t *szone, boolean_t verbose, region_t region, size_t bytes_at_end)
+{
+    unsigned           counts[1024];
+    unsigned           in_use = 0;
+    void               *start = SMALL_REGION_ADDRESS(region);
+    void               *limit = SMALL_REGION_END(region) - bytes_at_end;
+    msize_t            msize_and_free;
+    msize_t            msize;
+    unsigned           ci;
+    _SIMPLE_STRING     b;
+

     memset(counts, 0, 1024 * sizeof(unsigned));
     while (start < limit) {
     memset(counts, 0, 1024 * sizeof(unsigned));
     while (start < limit) {

-       msize_t msize_and_free = small_meta_header((void *)start)[0];
-       msize_t msize = msize_and_free & ~ SMALL_IS_FREE;
+       msize_and_free = *SMALL_METADATA_FOR_PTR(start);
+       msize = msize_and_free & ~ SMALL_IS_FREE;

        if (!(msize_and_free & SMALL_IS_FREE)) {
            // block in use
        if (!(msize_and_free & SMALL_IS_FREE)) {
            // block in use

-           if (msize < 1024) counts[msize]++;
+           if (msize < 1024)
+               counts[msize]++;

            in_use++;
        }
            in_use++;
        }

-       start += msize << SHIFT_SMALL_QUANTUM;
-    }
-    malloc_printf("Small region [%p-%p, %y]\tIn_use=%d ", SMALL_REGION_ADDRESS(*region), SMALL_REGION_END(*region), (int)SMALL_REGION_SIZE, in_use);
-    if (bytes_at_end) malloc_printf("Untouched=%y ", bytes_at_end);
-    if (verbose && in_use) {
-       unsigned        ci = 0;
-       malloc_printf("\n\tSizes in use: "); 
-       while (ci < 1024) {
-           if (counts[ci]) {
-               malloc_printf("%d[%d] ", ci << SHIFT_SMALL_QUANTUM, counts[ci]);
-           }
-           ci++;
+       start += SMALL_BYTES_FOR_MSIZE(msize);
+    }
+    if ((b = _simple_salloc()) != NULL) {
+       _simple_sprintf(b, "Small region [%p-%p, %y]\tIn_use=%d ",
+         SMALL_REGION_ADDRESS(region), SMALL_REGION_END(region), (int)SMALL_REGION_SIZE, in_use);
+       if (bytes_at_end)
+           _simple_sprintf(b, "Untouched=%ly", bytes_at_end);
+       if (verbose && in_use) {
+           _simple_sappend(b, "\n\tSizes in use: "); 
+           for (ci = 0; ci < 1024; ci++)
+               if (counts[ci])
+                   _simple_sprintf(b, "%d[%d] ", SMALL_BYTES_FOR_MSIZE(ci), counts[ci]);

        }
        }

+       _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX, "%s\n", _simple_string(b));
+       _simple_sfree(b);

     }
     }

-    malloc_printf("\n");

 }
 
 static boolean_t
 }
 
 static boolean_t

-small_free_list_check(szone_t *szone, grain_t grain) {
-    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
+small_free_list_check(szone_t *szone, grain_t grain)
+{

     unsigned   count = 0;
     free_list_t        *ptr = szone->small_free_list[grain];
     free_list_t        *previous = NULL;
     unsigned   count = 0;
     free_list_t        *ptr = szone->small_free_list[grain];
     free_list_t        *previous = NULL;

-    while (ptr) {
-       msize_t msize_and_free = small_meta_header(ptr)[0];
-       count++;
-       if (!(msize_and_free & SMALL_IS_FREE)) {
-           malloc_printf("*** malloc[%d]: In-use ptr in free list grain=%d count=%d ptr=%p\n", getpid(), grain, count, ptr);
-           return 0;
-       }
-       if (((unsigned)ptr) & (SMALL_QUANTUM - 1)) {
-           malloc_printf("*** malloc[%d]: Unaligned ptr in free list grain=%d  count=%d ptr=%p\n", getpid(), grain, count, ptr);
-           return 0;
-       }
-       if (!small_region_for_ptr_no_lock(szone, ptr)) {
-           malloc_printf("*** malloc[%d]: Ptr not in szone grain=%d  count=%d ptr=%p\n", getpid(), grain, count, ptr);
-           return 0;
-       }
-       free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
-       if (ptr->previous != previous) {
-           malloc_printf("*** malloc[%d]: Previous incorrectly set grain=%d  count=%d ptr=%p\n", getpid(), grain, count, ptr);
-           return 0;
-       }
-       previous = ptr;
-       ptr = ptr->next;
+    msize_t    msize_and_free;
+
+    CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
+    while (ptr) {
+        msize_and_free = *SMALL_METADATA_FOR_PTR(ptr);
+        count++;
+        if (!(msize_and_free & SMALL_IS_FREE)) {
+            malloc_printf("*** in-use ptr in free list grain=%d count=%d ptr=%p\n", grain, count, ptr);
+            return 0;
+        }
+        if (((uintptr_t)ptr) & (SMALL_QUANTUM - 1)) {
+            malloc_printf("*** unaligned ptr in free list grain=%d  count=%d ptr=%p\n", grain, count, ptr);
+            return 0;
+        }
+        if (!small_region_for_ptr_no_lock(szone, ptr)) {
+            malloc_printf("*** ptr not in szone grain=%d  count=%d ptr=%p\n", grain, count, ptr);
+            return 0;
+        }
+        free_list_checksum(szone, ptr, __PRETTY_FUNCTION__);
+        if (free_list_unchecksum_ptr(ptr->previous) != previous) {
+            malloc_printf("*** previous incorrectly set grain=%d  count=%d ptr=%p\n", grain, count, ptr);
+            return 0;
+        }
+        previous = ptr;
+        ptr = free_list_unchecksum_ptr(ptr->next);

     }
     return 1;
 }
 
     }
     return 1;
 }
 

-/********************* LARGE ENTRY UTILITIES   ************************/
+/*******************************************************************************
+ * Region hash implementation
+ *
+ * This is essentially a duplicate of the existing Large allocator hash, minus
+ * the ability to remove entries.  The two should be combined eventually.
+ ******************************************************************************/
+#pragma mark region hash
+
+/*
+ * hash_lookup_region_no_lock - Scan a hash ring looking for an entry for a 
+ * given region.
+ *
+ * FIXME: If consecutive queries of the same region are likely, a one-entry
+ * cache would likely be a significant performance win here.
+ */
+static region_t *
+hash_lookup_region_no_lock(region_t *regions, size_t num_entries, region_t r) {
+  size_t index, hash_index;
+  region_t *entry;
+
+  if (!num_entries)
+    return 0;
+  
+  index = hash_index = ((uintptr_t)r >> 20) % num_entries;
+  do {
+    entry = regions + index;
+    if (*entry == 0)
+      return 0;
+    if (*entry == r)
+      return entry;
+    if (++index == num_entries)
+      index = 0;
+  } while (index != hash_index);
+  return 0;
+}
+
+/*
+ * hash_region_insert_no_lock - Insert a region into the hash ring.
+ */
+static void
+hash_region_insert_no_lock(region_t *regions, size_t num_entries, region_t r) {
+  size_t index, hash_index;
+  region_t *entry;
+  
+  index = hash_index = ((uintptr_t)r >> 20) % num_entries;
+  do {
+    entry = regions + index;
+    if (*entry == 0) {
+           *entry = r;
+      return;
+    }
+    if (++index == num_entries)
+           index = 0;
+  } while (index != hash_index);
+}
+
+/*
+ * hash_regions_alloc_no_lock - Allocate space for a number of entries.  This
+ * must be a VM allocation as to avoid recursing between allocating a new small
+ * region, and asking the small region to allocate space for the new list of
+ * regions.
+ */
+static region_t *
+hash_regions_alloc_no_lock(szone_t *szone, size_t num_entries)
+{
+  size_t size = num_entries * sizeof(region_t);
+       return allocate_pages(szone, round_page(size), 0, 0, VM_MEMORY_MALLOC);
+}
+
+/*
+ * hash_regions_grow_no_lock - Grow the hash ring, and rehash the entries.
+ * Return the new region and new size to update the szone.  Do not deallocate
+ * the old entries since someone may still be allocating them.
+ */
+static region_t *
+hash_regions_grow_no_lock(szone_t *szone, region_t *regions, size_t old_size,
+                          size_t *new_size)
+{
+  // double in size and allocate memory for the regions
+  *new_size = old_size * 2 + 1;
+  region_t *new_regions = hash_regions_alloc_no_lock(szone, *new_size);
+
+  // rehash the entries into the new list
+  size_t index;
+  for (index = 0; index < old_size; ++index) {
+    region_t r = regions[index];
+    if (r != 0)
+      hash_region_insert_no_lock(new_regions, *new_size, r);
+  }
+  return new_regions;
+}
+
+/*******************************************************************************
+ * Large allocator implementation
+ ******************************************************************************/
+#pragma mark large allocator

 
 #if DEBUG_MALLOC
 
 static void
 
 #if DEBUG_MALLOC
 
 static void

-large_debug_print(szone_t *szone) {
-    unsigned   num_large_entries = szone->num_large_entries;
-    unsigned   index = num_large_entries;
-    while (index--) {
-       large_entry_t   *range = szone->large_entries + index;
-       large_entry_t   entry = *range;
-       if (!LARGE_ENTRY_IS_EMPTY(entry)) {
-           malloc_printf("%d: %p(%y);  ", index, LARGE_ENTRY_ADDRESS(entry), LARGE_ENTRY_SIZE(entry));
-       }
+large_debug_print(szone_t *szone)
+{
+    unsigned           num_large_entries = szone->num_large_entries;
+    unsigned           index = num_large_entries;
+    large_entry_t      *range;
+    _SIMPLE_STRING     b = _simple_salloc();
+
+    if (b) {
+       for (index = 0, range = szone->large_entries; index < szone->num_large_entries; index++, range++)
+           if (!LARGE_ENTRY_IS_EMPTY(*range))
+               _simple_sprintf(b, "%d: %p(%y);  ", index, LARGE_ENTRY_ADDRESS(*range), LARGE_ENTRY_SIZE(*range));
+
+       _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX, "%s\n", _simple_string(b));
+       _simple_sfree(b);

     }
     }

-    malloc_printf("\n");

 }
 #endif
 
 }
 #endif
 

+/*
+ * Scan the hash ring looking for an entry for the given pointer.
+ */

 static large_entry_t *
 static large_entry_t *

-large_entry_for_pointer_no_lock(szone_t *szone,
-  const void *ptr) {
-    // result only valid during a lock
-    unsigned   num_large_entries = szone->num_large_entries;
-    unsigned   hash_index;
-    unsigned   index;
-    if (!num_large_entries) return NULL;
-    hash_index = ((unsigned)ptr >> vm_page_shift) % num_large_entries;
+large_entry_for_pointer_no_lock(szone_t *szone, const void *ptr)
+{
+    // result only valid with lock held
+    unsigned           num_large_entries = szone->num_large_entries;
+    unsigned           hash_index;
+    unsigned           index;
+    large_entry_t      *range;
+
+    if (!num_large_entries)
+       return NULL;
+    hash_index = ((uintptr_t)ptr >> vm_page_shift) % num_large_entries;

     index = hash_index;
     do {
     index = hash_index;
     do {

-       large_entry_t   *range = szone->large_entries + index;
-       large_entry_t   entry = *range;
-       if (LARGE_ENTRY_MATCHES(entry, ptr)) return range;
-       if (LARGE_ENTRY_IS_EMPTY(entry)) return NULL; // end of chain
-       index++; if (index == num_large_entries) index = 0;
+       range = szone->large_entries + index;
+       if (LARGE_ENTRY_MATCHES(*range, ptr))
+           return range;
+       if (LARGE_ENTRY_IS_EMPTY(*range))
+           return NULL; // end of chain
+       index++;
+       if (index == num_large_entries)
+           index = 0;

     } while (index != hash_index);
     return NULL;
 }
 
 static void
     } while (index != hash_index);
     return NULL;
 }
 
 static void

-large_entry_insert_no_lock(szone_t *szone, large_entry_t range) {
-    unsigned   num_large_entries = szone->num_large_entries;
-    unsigned   hash_index = (range.address_and_num_pages >> vm_page_shift)
-      % num_large_entries;
-    unsigned   index = hash_index;
-//    malloc_printf("Before insertion of %p\n", LARGE_ENTRY_ADDRESS(range));
+large_entry_insert_no_lock(szone_t *szone, large_entry_t range)
+{
+    unsigned           num_large_entries = szone->num_large_entries;
+    unsigned           hash_index = (range.address_and_num_pages >> vm_page_shift) % num_large_entries;
+    unsigned           index = hash_index;
+    large_entry_t      *entry;
+

     do {
     do {

-       large_entry_t   *entry = szone->large_entries + index;
+       entry = szone->large_entries + index;

        if (LARGE_ENTRY_IS_EMPTY(*entry)) {
            *entry = range;
            return; // end of chain
        }
        if (LARGE_ENTRY_IS_EMPTY(*entry)) {
            *entry = range;
            return; // end of chain
        }

-       index++; if (index == num_large_entries) index = 0;
+       index++;
+       if (index == num_large_entries)
+           index = 0;

     } while (index != hash_index);
 }
 
 static INLINE void
     } while (index != hash_index);
 }
 
 static INLINE void

-large_entries_rehash_after_entry_no_lock(szone_t *szone,
-  large_entry_t *entry) {
-    unsigned   num_large_entries = szone->num_large_entries;
-    unsigned   hash_index = entry - szone->large_entries;
-    unsigned   index = hash_index;
+large_entries_rehash_after_entry_no_lock(szone_t *szone, large_entry_t *entry)
+{
+    unsigned           num_large_entries = szone->num_large_entries;
+    unsigned           hash_index = entry - szone->large_entries;
+    unsigned           index = hash_index;
+    large_entry_t      range;
+

     do {
     do {

-       large_entry_t   range;
-       index++; if (index == num_large_entries) index = 0;
+       index++;
+       if (index == num_large_entries)
+           index = 0;

        range = szone->large_entries[index];
        range = szone->large_entries[index];

-       if (LARGE_ENTRY_IS_EMPTY(range)) return;
+       if (LARGE_ENTRY_IS_EMPTY(range))
+           return;

        szone->large_entries[index].address_and_num_pages = 0;
        large_entry_insert_no_lock(szone, range); // this will reinsert in the
                                                  // proper place
        szone->large_entries[index].address_and_num_pages = 0;
        large_entry_insert_no_lock(szone, range); // this will reinsert in the
                                                  // proper place


@@ -2015,55 +2922,66 @@ large_entries_rehash_after_entry_no_lock(szone_t *szone,
 }
 
 static INLINE large_entry_t *
 }
 
 static INLINE large_entry_t *

-large_entries_alloc_no_lock(szone_t *szone,
-  unsigned num) {
+large_entries_alloc_no_lock(szone_t *szone, unsigned num)
+{

     size_t     size = num * sizeof(large_entry_t);
     boolean_t  is_vm_allocation = size >= LARGE_THRESHOLD;
     size_t     size = num * sizeof(large_entry_t);
     boolean_t  is_vm_allocation = size >= LARGE_THRESHOLD;

+

     if (is_vm_allocation) {
        // Note that we allocate memory (via a system call) under a spin lock
        // That is certainly evil, however it's very rare in the lifetime of a process
        // The alternative would slow down the normal case
     if (is_vm_allocation) {
        // Note that we allocate memory (via a system call) under a spin lock
        // That is certainly evil, however it's very rare in the lifetime of a process
        // The alternative would slow down the normal case

-       return (void *)allocate_pages(szone, round_page(size), 0, 0, VM_MAKE_TAG(VM_MEMORY_MALLOC_LARGE));
+       return allocate_pages(szone, round_page(size), 0, 0, VM_MEMORY_MALLOC_LARGE);

     } else {
     } else {

-       return small_malloc_cleared_no_lock(szone, (size + SMALL_QUANTUM - 1) >> SHIFT_SMALL_QUANTUM);
+       return small_malloc_cleared_no_lock(szone, SMALL_MSIZE_FOR_BYTES(size + SMALL_QUANTUM - 1));

     }
 }
 
 static void
     }
 }
 
 static void

-large_entries_free_no_lock(szone_t *szone, large_entry_t *entries, unsigned num, vm_range_t *range_to_deallocate) {
+large_entries_free_no_lock(szone_t *szone, large_entry_t *entries, unsigned num, vm_range_t *range_to_deallocate)
+{

     // returns range to deallocate
     // returns range to deallocate

-    size_t     size = num * sizeof(large_entry_t);
-    boolean_t  is_vm_allocation = size >= LARGE_THRESHOLD;
-    // malloc_printf("In large_entries_free_no_lock %d %d\n", num, is_vm_allocation);
+    size_t             size = num * sizeof(large_entry_t);
+    boolean_t          is_vm_allocation = size >= LARGE_THRESHOLD;
+    region_t   *region;
+    msize_t            msize_and_free;
+    

     if (is_vm_allocation) {
        range_to_deallocate->address = (vm_address_t)entries;
        range_to_deallocate->size = round_page(size);
     } else {
        range_to_deallocate->size = 0;
     if (is_vm_allocation) {
        range_to_deallocate->address = (vm_address_t)entries;
        range_to_deallocate->size = round_page(size);
     } else {
        range_to_deallocate->size = 0;

-       small_region_t  *region = small_region_for_ptr_no_lock(szone, entries);
-       msize_t         msize_and_free = small_meta_header(entries)[0];
+       region = small_region_for_ptr_no_lock(szone, entries);
+       msize_and_free = *SMALL_METADATA_FOR_PTR(entries);

        if (msize_and_free & SMALL_IS_FREE) {
        if (msize_and_free & SMALL_IS_FREE) {

-           szone_error(szone, "Object already freed being freed", entries);
+           szone_error(szone, "object already freed being freed", entries, NULL);

            return;
        }
        small_free_no_lock(szone, region, entries, msize_and_free);
     }
 }
 
            return;
        }
        small_free_no_lock(szone, region, entries, msize_and_free);
     }
 }
 

-static void
-large_entries_grow_no_lock(szone_t *szone, vm_range_t *range_to_deallocate) {
+static large_entry_t *
+large_entries_grow_no_lock(szone_t *szone, vm_range_t *range_to_deallocate)
+{

     // sets range_to_deallocate
     unsigned           old_num_entries = szone->num_large_entries;
     large_entry_t      *old_entries = szone->large_entries;
     // sets range_to_deallocate
     unsigned           old_num_entries = szone->num_large_entries;
     large_entry_t      *old_entries = szone->large_entries;

-    unsigned           new_num_entries = (old_num_entries) ? old_num_entries
-      * 2 + 1 : 63; // always an odd number for good hashing
+    unsigned           new_num_entries = (old_num_entries) ? old_num_entries * 2 + 1 : 63; // always an odd number for good hashing

     large_entry_t      *new_entries = large_entries_alloc_no_lock(szone, new_num_entries);
     unsigned           index = old_num_entries;
     large_entry_t      *new_entries = large_entries_alloc_no_lock(szone, new_num_entries);
     unsigned           index = old_num_entries;

+    large_entry_t      oldRange;
+
+    // if the allocation of new entries failed, bail
+       if (new_entries == NULL)
+               return NULL;
+

     szone->num_large_entries = new_num_entries;
     szone->large_entries = new_entries;
     szone->num_large_entries = new_num_entries;
     szone->large_entries = new_entries;

-    // malloc_printf("_grow_large_entries old_num_entries=%d new_num_entries=%d %p\n", old_num_entries, new_num_entries, old_entries);
+
+    /* rehash entries into the new list */

     while (index--) {
     while (index--) {

-       large_entry_t   oldRange = old_entries[index];
+       oldRange = old_entries[index];

        if (!LARGE_ENTRY_IS_EMPTY(oldRange)) {
            large_entry_insert_no_lock(szone, oldRange);
        }
        if (!LARGE_ENTRY_IS_EMPTY(oldRange)) {
            large_entry_insert_no_lock(szone, oldRange);
        }


@@ -2073,33 +2991,31 @@ large_entries_grow_no_lock(szone_t *szone, vm_range_t *range_to_deallocate) {
     } else {
        range_to_deallocate->size = 0;
     }
     } else {
        range_to_deallocate->size = 0;
     }

+    return new_entries;

 }
 
 }
 

+// frees the specific entry in the size table
+// returns a range to truly deallocate

 static vm_range_t
 static vm_range_t

-large_free_no_lock(szone_t *szone, large_entry_t *entry) {
-    // frees the specific entry in the size table
-    // returns a range to truly deallocate
+large_free_no_lock(szone_t *szone, large_entry_t *entry)
+{

     vm_range_t         range;
     vm_range_t         range;

-    range.address = LARGE_ENTRY_ADDRESS(*entry);
-    range.size = LARGE_ENTRY_SIZE(*entry);
-    szone->num_large_objects_in_use --;
+    
+    range.address = (vm_address_t)LARGE_ENTRY_ADDRESS(*entry);
+    range.size = (vm_size_t)LARGE_ENTRY_SIZE(*entry);
+    szone->num_large_objects_in_use--;

     szone->num_bytes_in_large_objects -= range.size;
     if (szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) {
     szone->num_bytes_in_large_objects -= range.size;
     if (szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) {

-       protect(szone, range.address, range.size, VM_PROT_READ | VM_PROT_WRITE,
-         szone->debug_flags);
-       range.address -= 1 << vm_page_shift;
-       range.size += 2 * (1 << vm_page_shift);
-    }
-//    malloc_printf("Entry is %p=%d; cache is %p ; found=%p\n", entry,
-//      entry-szone->large_entries, szone->large_entries,
-//      large_entry_for_pointer_no_lock(szone, (void *)range.address));
+       protect((void *)range.address, range.size, VM_PROT_READ | VM_PROT_WRITE, szone->debug_flags);
+       range.address -= vm_page_size;
+       range.size += 2 * vm_page_size;
+    }

     entry->address_and_num_pages = 0;
     large_entries_rehash_after_entry_no_lock(szone, entry);
 #if DEBUG_MALLOC
     if (large_entry_for_pointer_no_lock(szone, (void *)range.address)) {
     entry->address_and_num_pages = 0;
     large_entries_rehash_after_entry_no_lock(szone, entry);
 #if DEBUG_MALLOC
     if (large_entry_for_pointer_no_lock(szone, (void *)range.address)) {

-       malloc_printf("*** malloc[%d]: Freed entry %p still in use; "
-         "num_large_entries=%d\n", getpid(), range.address,
-         szone->num_large_entries);
+       malloc_printf("*** freed entry %p still in use; num_large_entries=%d\n",
+         range.address, szone->num_large_entries);

        large_debug_print(szone);
        szone_sleep();
     }
        large_debug_print(szone);
        szone_sleep();
     }


@@ -2107,102 +3023,40 @@ large_free_no_lock(szone_t *szone, large_entry_t *entry) {
     return range;
 }
 
     return range;
 }
 

-static INLINE boolean_t
-try_realloc_small_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) {
-    // returns 1 on success
-    msize_t    *meta_headers = small_meta_headers(ptr);
-    msize_t    index = small_meta_index(ptr);
-    msize_t    old_msize = old_size >> SHIFT_SMALL_QUANTUM;
-    msize_t    new_msize = (new_size + SMALL_QUANTUM - 1) >> SHIFT_SMALL_QUANTUM;
-    void       *next_block = (char *)ptr + old_size;
-    msize_t    next_index = index + old_msize;
-    if (next_index >= NUM_SMALL_BLOCKS) {
-       // malloc_printf("try_realloc_small_in_place can't take place at end %p %d %d %d\n", ptr, old_size, new_size, next_index);
-       return 0;
-    }
-#if DEBUG_MALLOC
-    if ((vm_address_t)next_block & (SMALL_QUANTUM - 1)) {
-       szone_error(szone, "Internal invariant broken in realloc(next_block)", next_block);
-    }
-    if (meta_headers[index] != old_msize) malloc_printf("*** try_realloc_small_in_place incorrect old %d %d\n", meta_headers[index], old_msize);
-#endif
-    SZONE_LOCK(szone);
-    // If the next block is free, we coalesce
-    msize_t    next_msize_and_free;
-    msize_t    next_msize;
-    next_msize_and_free = meta_headers[next_index];
-    next_msize = next_msize_and_free & ~ SMALL_IS_FREE;
-    if (!(next_msize_and_free & SMALL_IS_FREE) || (old_msize + next_msize < new_msize)) {
-       SZONE_UNLOCK(szone);
-       return 0;
-    }
-    // malloc_printf("Small realloc in place for %p;  current msize=%db(%d) next=%p next_msize=%d wanted=%db(%d)\n", ptr, old_size, meta_headers[index], next_block, next_msize, new_size, new_msize);
-    small_free_list_remove_ptr(szone, next_block, next_msize);
-    small_meta_header_set_middle(meta_headers, next_index);
-    msize_t    leftover_msize = old_msize + next_msize - new_msize;
-    if (leftover_msize) {
-       void    *leftover = ptr + (new_msize << SHIFT_SMALL_QUANTUM);
-       // malloc_printf("Leftover in realloc in place %p msize=%d\n", leftover, leftover_msize);
-       small_free_list_add_ptr(szone, leftover, leftover_msize);
-       msize_t leftover_index = index + new_msize;
-       small_meta_header_set_is_free(meta_headers, leftover_index, leftover_msize);
-    }
-#if DEBUG_MALLOC
-    if ((new_msize << SHIFT_SMALL_QUANTUM) >= LARGE_THRESHOLD) {
-       malloc_printf("*** Realloc in place for %p exceeded msize=%d\n", new_msize);
-    }
-#endif
-    small_meta_header_set_in_use(meta_headers, index, new_msize);
-#if DEBUG_MALLOC
-    if (LOG(szone,ptr)) {
-       malloc_printf("In szone_realloc(), ptr=%p, msize=%d\n", ptr, small_meta_header(ptr)[0]);
-    }
-#endif
-    szone->num_bytes_in_small_objects += (new_msize - old_msize) << SHIFT_SMALL_QUANTUM;
-    SZONE_UNLOCK(szone);
-    CHECK(szone, __PRETTY_FUNCTION__);
-//    malloc_printf("Extended ptr %p for realloc old=%d desired=%d new=%d "
-//      "leftover=%d\n", ptr, (unsigned)old_size, (unsigned)new_size,
-//      (unsigned)szone_size(szone, ptr), leftover_msize << SHIFT_SMALL_QUANTUM);
-    return 1;
-}
-

 static kern_return_t
 static kern_return_t

-large_in_use_enumerator(task_t task, void *context,
-  unsigned type_mask, vm_address_t large_entries_address, unsigned num_entries,
-  memory_reader_t reader, vm_range_recorder_t recorder) {
+large_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t large_entries_address, unsigned num_entries, memory_reader_t reader, vm_range_recorder_t recorder)
+{

     unsigned           index = 0;
     vm_range_t         buffer[MAX_RECORDER_BUFFER];
     unsigned           count = 0;
     large_entry_t      *entries;
     kern_return_t      err;
     unsigned           index = 0;
     vm_range_t         buffer[MAX_RECORDER_BUFFER];
     unsigned           count = 0;
     large_entry_t      *entries;
     kern_return_t      err;

-    err = reader(task, large_entries_address,
-      sizeof(large_entry_t) * num_entries, (void **)&entries);
-    if (err) return err;
+    vm_range_t         range;
+    large_entry_t      entry;
+
+    err = reader(task, large_entries_address, sizeof(large_entry_t) * num_entries, (void **)&entries);
+    if (err)
+       return err;

     index = num_entries;
     index = num_entries;

-    if ((type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE)
-      && (num_entries * sizeof(large_entry_t) >= LARGE_THRESHOLD)) {
-       vm_range_t      range;
+    if ((type_mask & MALLOC_ADMIN_REGION_RANGE_TYPE) &&
+      (num_entries * sizeof(large_entry_t) >= LARGE_THRESHOLD)) {

        range.address = large_entries_address;
        range.size = round_page(num_entries * sizeof(large_entry_t));
        recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &range, 1);
     }
        range.address = large_entries_address;
        range.size = round_page(num_entries * sizeof(large_entry_t));
        recorder(task, context, MALLOC_ADMIN_REGION_RANGE_TYPE, &range, 1);
     }

-    if (type_mask & (MALLOC_PTR_IN_USE_RANGE_TYPE
-      | MALLOC_PTR_REGION_RANGE_TYPE))
-      while (index--) {
-       large_entry_t   entry = entries[index];
-       if (!LARGE_ENTRY_IS_EMPTY(entry)) {
-           vm_range_t  range;
-           range.address = LARGE_ENTRY_ADDRESS(entry);
-           range.size = LARGE_ENTRY_SIZE(entry);
-           buffer[count++] = range;
-           if (count >= MAX_RECORDER_BUFFER) {
-               recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE
-                 | MALLOC_PTR_REGION_RANGE_TYPE, buffer, count);
-               count = 0;
+    if (type_mask & (MALLOC_PTR_IN_USE_RANGE_TYPE | MALLOC_PTR_REGION_RANGE_TYPE))
+       while (index--) {
+           entry = entries[index];
+           if (!LARGE_ENTRY_IS_EMPTY(entry)) {
+               range.address = (vm_address_t)LARGE_ENTRY_ADDRESS(entry);
+               range.size = (vm_size_t)LARGE_ENTRY_SIZE(entry);
+               buffer[count++] = range;
+               if (count >= MAX_RECORDER_BUFFER) {
+                   recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE | MALLOC_PTR_REGION_RANGE_TYPE, buffer, count);
+                   count = 0;
+               }

            }
        }
            }
        }

-    }

     if (count) {
        recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE
          | MALLOC_PTR_REGION_RANGE_TYPE, buffer, count);
     if (count) {
        recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE
          | MALLOC_PTR_REGION_RANGE_TYPE, buffer, count);


@@ -2213,36 +3067,48 @@ large_in_use_enumerator(task_t task, void *context,
 /********************* HUGE ENTRY UTILITIES    ************************/
 
 static huge_entry_t *
 /********************* HUGE ENTRY UTILITIES    ************************/
 
 static huge_entry_t *

-huge_entry_for_pointer_no_lock(szone_t *szone,
-  const void *ptr) {
-    unsigned   index = szone->num_huge_entries;
-    while (index--) {
-       huge_entry_t    *huge = szone->huge_entries + index;
-       if (huge->address == (vm_address_t)ptr) return huge;
+huge_entry_for_pointer_no_lock(szone_t *szone, const void *ptr)
+{
+    unsigned           index;
+    huge_entry_t       *huge;
+
+    for (index = szone->num_huge_entries, huge = szone->huge_entries;
+        index > 0;
+        index--, huge++) {
+    
+       if ((void *)huge->address == ptr)
+           return huge;

     }
     return NULL;
 }
 
 static boolean_t
     }
     return NULL;
 }
 
 static boolean_t

-huge_entry_append(szone_t *szone, huge_entry_t huge) {
+huge_entry_append(szone_t *szone, huge_entry_t huge)
+{
+    huge_entry_t       *new_huge_entries = NULL, *old_huge_entries;
+    unsigned           num_huge_entries;
+    

     // We do a little dance with locking because doing allocation (even in the
     // default szone) may cause something to get freed in this szone, with a
     // deadlock
     // Returns 1 on success
     // We do a little dance with locking because doing allocation (even in the
     // default szone) may cause something to get freed in this szone, with a
     // deadlock
     // Returns 1 on success

-    huge_entry_t       *new_huge_entries = NULL;

     SZONE_LOCK(szone);
     SZONE_LOCK(szone);

-    while (1) {
-       unsigned        num_huge_entries;
+    for (;;) {

        num_huge_entries = szone->num_huge_entries;
        SZONE_UNLOCK(szone);
        num_huge_entries = szone->num_huge_entries;
        SZONE_UNLOCK(szone);

-//     malloc_printf("In huge_entry_append currentEntries=%d\n", num_huge_entries);
-       if (new_huge_entries) szone_free(szone, new_huge_entries);
+       /* check for counter wrap */
+       if ((num_huge_entries + 1) < num_huge_entries)
+               return 0;
+       /* stale allocation from last time around the loop? */
+       if (new_huge_entries)
+           szone_free(szone, new_huge_entries);

        new_huge_entries = szone_malloc(szone, (num_huge_entries + 1) * sizeof(huge_entry_t));
        new_huge_entries = szone_malloc(szone, (num_huge_entries + 1) * sizeof(huge_entry_t));

-       if (new_huge_entries == NULL) return 0;
+       if (new_huge_entries == NULL)
+           return 0;

        SZONE_LOCK(szone);
        if (num_huge_entries == szone->num_huge_entries) {
            // No change - our malloc still applies
        SZONE_LOCK(szone);
        if (num_huge_entries == szone->num_huge_entries) {
            // No change - our malloc still applies

-           huge_entry_t        *old_huge_entries = szone->huge_entries;
+           old_huge_entries = szone->huge_entries;

            if (num_huge_entries) {
                memcpy(new_huge_entries, old_huge_entries, num_huge_entries * sizeof(huge_entry_t));
            }
            if (num_huge_entries) {
                memcpy(new_huge_entries, old_huge_entries, num_huge_entries * sizeof(huge_entry_t));
            }


@@ -2250,7 +3116,6 @@ huge_entry_append(szone_t *szone, huge_entry_t huge) {
            szone->huge_entries = new_huge_entries;
            SZONE_UNLOCK(szone);
            szone_free(szone, old_huge_entries);
            szone->huge_entries = new_huge_entries;
            SZONE_UNLOCK(szone);
            szone_free(szone, old_huge_entries);

-//         malloc_printf("Done huge_entry_append now=%d\n", szone->num_huge_entries);

            return 1;
        }
        // try again!
            return 1;
        }
        // try again!


@@ -2258,53 +3123,58 @@ huge_entry_append(szone_t *szone, huge_entry_t huge) {
 }
 
 static kern_return_t
 }
 
 static kern_return_t

-huge_in_use_enumerator(task_t task, void *context,
-  unsigned type_mask, vm_address_t huge_entries_address, unsigned num_entries,
-  memory_reader_t reader, vm_range_recorder_t recorder) {
+huge_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t huge_entries_address, unsigned num_entries, memory_reader_t reader, vm_range_recorder_t recorder)
+{

     huge_entry_t       *entries;
     kern_return_t      err;
     huge_entry_t       *entries;
     kern_return_t      err;

-    err = reader(task, huge_entries_address, sizeof(huge_entry_t) * num_entries,
-      (void **)&entries);
-    if (err) return err;
-    if (num_entries) {
-       recorder(task, context,
-         MALLOC_PTR_IN_USE_RANGE_TYPE | MALLOC_PTR_REGION_RANGE_TYPE, entries,
-         num_entries);
-    }
+    
+    err = reader(task, huge_entries_address, sizeof(huge_entry_t) * num_entries, (void **)&entries);
+    if (err)
+       return err;
+    if (num_entries)
+       recorder(task, context, MALLOC_PTR_IN_USE_RANGE_TYPE | MALLOC_PTR_REGION_RANGE_TYPE, entries, num_entries);
+

     return 0;
 }
 
 static void *
     return 0;
 }
 
 static void *

-large_and_huge_malloc(szone_t *szone, unsigned num_pages) {
-    vm_address_t       addr = 0;
+large_and_huge_malloc(szone_t *szone, size_t num_pages)
+{
+    void               *addr;

     vm_range_t         range_to_deallocate;
     vm_range_t         range_to_deallocate;

-    if (!num_pages) num_pages = 1; // minimal allocation size for this szone
-//    malloc_printf("In large_and_huge_malloc for %y\n", num_pages * (1 << vm_page_shift));
+    huge_entry_t       huge_entry;
+    size_t             size;
+    large_entry_t      large_entry;
+    
+    if (!num_pages)
+           num_pages = 1; // minimal allocation size for this szone
+    size = (size_t)num_pages << vm_page_shift;

     range_to_deallocate.size = 0;
     if (num_pages >= (1 << vm_page_shift)) {
     range_to_deallocate.size = 0;
     if (num_pages >= (1 << vm_page_shift)) {

-       huge_entry_t    huge;
-       huge.size = num_pages << vm_page_shift;
-       addr = allocate_pages(szone, huge.size, 0, szone->debug_flags, VM_MAKE_TAG(VM_MEMORY_MALLOC_HUGE));
-       if (!addr) return NULL;
-       huge.address = addr;
-       if (! huge_entry_append(szone, huge)) return NULL;
+       addr = allocate_pages(szone, size, 0, szone->debug_flags, VM_MEMORY_MALLOC_HUGE);
+       if (addr == NULL)
+           return NULL;
+       huge_entry.size = size;
+       huge_entry.address = (vm_address_t)addr;
+       if (!huge_entry_append(szone, huge_entry))
+           return NULL;        // we are leaking the allocation here

        SZONE_LOCK(szone);
        SZONE_LOCK(szone);

-       szone->num_bytes_in_huge_objects += huge.size;
+       szone->num_bytes_in_huge_objects += size;

     } else {
     } else {

-       vm_size_t               size = num_pages << vm_page_shift;
-       large_entry_t           entry;
-       addr = allocate_pages(szone, size, 0, szone->debug_flags, VM_MAKE_TAG(VM_MEMORY_MALLOC_LARGE));
+
+       addr = allocate_pages(szone, size, 0, szone->debug_flags, VM_MEMORY_MALLOC_LARGE);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (LOG(szone, addr)) malloc_printf("In szone_malloc true large allocation at %p for %y\n", (void *)addr, size);
+       if (LOG(szone, addr))
+           malloc_printf("in szone_malloc true large allocation at %p for %ly\n", (void *)addr, size);

 #endif
        SZONE_LOCK(szone);
 #endif
        SZONE_LOCK(szone);

-       if (!addr) {
+       if (addr == NULL) {

            SZONE_UNLOCK(szone);
            return NULL;
        }
 #if DEBUG_MALLOC
            SZONE_UNLOCK(szone);
            return NULL;
        }
 #if DEBUG_MALLOC

-       if (large_entry_for_pointer_no_lock(szone, (void *)addr)) {
-           malloc_printf("Freshly allocated is already in use: %p\n", addr);
+       if (large_entry_for_pointer_no_lock(szone, addr)) {
+           malloc_printf("freshly allocated is already in use: %p\n", addr);

            large_debug_print(szone);
            szone_sleep();
        }
            large_debug_print(szone);
            szone_sleep();
        }


@@ -2312,54 +3182,53 @@ large_and_huge_malloc(szone_t *szone, unsigned num_pages) {
        if ((szone->num_large_objects_in_use + 1) * 4 > szone->num_large_entries) {
            // density of hash table too high; grow table
            // we do that under lock to avoid a race
        if ((szone->num_large_objects_in_use + 1) * 4 > szone->num_large_entries) {
            // density of hash table too high; grow table
            // we do that under lock to avoid a race

-           // malloc_printf("In szone_malloc growing hash table current=%d\n", szone->num_large_entries);
-           large_entries_grow_no_lock(szone, &range_to_deallocate);
+           large_entry_t *entries = large_entries_grow_no_lock(szone, &range_to_deallocate);
+           if (entries == NULL) {
+               SZONE_UNLOCK(szone);
+               return NULL;
+           }

        }
        }

-//     malloc_printf("Inserting large entry (%p, %y)\n", addr, num_pages * (1 << vm_page_shift));
-       entry.address_and_num_pages = addr | num_pages;
+       large_entry.address_and_num_pages = (uintptr_t)addr | num_pages;

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (large_entry_for_pointer_no_lock(szone, (void *)addr)) {
-           malloc_printf("Entry about to be added already in use: %p\n",
-             addr);
+       if (large_entry_for_pointer_no_lock(szone, addr)) {
+           malloc_printf("entry about to be added already in use: %p\n", addr);

            large_debug_print(szone);
            szone_sleep();
        }
 #endif
            large_debug_print(szone);
            szone_sleep();
        }
 #endif

-       large_entry_insert_no_lock(szone, entry);
+       large_entry_insert_no_lock(szone, large_entry);

 #if DEBUG_MALLOC
        if (!large_entry_for_pointer_no_lock(szone, (void *)addr)) {
 #if DEBUG_MALLOC
        if (!large_entry_for_pointer_no_lock(szone, (void *)addr)) {

-           malloc_printf("Can't find entry just added\n");
+           malloc_printf("can't find entry just added\n");

            large_debug_print(szone);
            szone_sleep();
        }
 #endif
            large_debug_print(szone);
            szone_sleep();
        }
 #endif

-//     malloc_printf("Inserted large entry (%p, %d pages)\n", addr,
-//       num_pages);

        szone->num_large_objects_in_use ++;
        szone->num_bytes_in_large_objects += size;
     }
     SZONE_UNLOCK(szone);
     if (range_to_deallocate.size) {
        szone->num_large_objects_in_use ++;
        szone->num_bytes_in_large_objects += size;
     }
     SZONE_UNLOCK(szone);
     if (range_to_deallocate.size) {

-       deallocate_pages(szone, range_to_deallocate.address, range_to_deallocate.size, 0); // we deallocate outside the lock
-       // malloc_printf("Deallocated large entries %d\n", range_to_deallocate.size);
+       deallocate_pages(szone, (void *)range_to_deallocate.address, range_to_deallocate.size, 0); // we deallocate outside the lock

     }
     return (void *)addr;
 }
 
 static INLINE void
     }
     return (void *)addr;
 }
 
 static INLINE void

-free_large_or_huge(szone_t *szone, void *ptr) {
+free_large_or_huge(szone_t *szone, void *ptr)
+{

     // We have established ptr is page-aligned and not tiny nor small
     large_entry_t      *entry;
     vm_range_t         vm_range_to_deallocate;
     huge_entry_t       *huge;
     // We have established ptr is page-aligned and not tiny nor small
     large_entry_t      *entry;
     vm_range_t         vm_range_to_deallocate;
     huge_entry_t       *huge;

+    

     SZONE_LOCK(szone);
     entry = large_entry_for_pointer_no_lock(szone, ptr);
     if (entry) {
     SZONE_LOCK(szone);
     entry = large_entry_for_pointer_no_lock(szone, ptr);
     if (entry) {

-//     malloc_printf("Ready for deallocation [%p-%y]\n", LARGE_ENTRY_ADDRESS(*entry), LARGE_ENTRY_SIZE(*entry));

        vm_range_to_deallocate = large_free_no_lock(szone, entry);
 #if DEBUG_MALLOC
        if (large_entry_for_pointer_no_lock(szone, ptr)) {
        vm_range_to_deallocate = large_free_no_lock(szone, entry);
 #if DEBUG_MALLOC
        if (large_entry_for_pointer_no_lock(szone, ptr)) {

-           malloc_printf("*** malloc[%d]: Just after freeing %p still in use num_large_entries=%d\n", getpid(), ptr, szone->num_large_entries);
+           malloc_printf("*** just after freeing %p still in use num_large_entries=%d\n", ptr, szone->num_large_entries);

            large_debug_print(szone);
            szone_sleep();
        }
            large_debug_print(szone);
            szone_sleep();
        }


@@ -2367,50 +3236,49 @@ free_large_or_huge(szone_t *szone, void *ptr) {
     } else if ((huge = huge_entry_for_pointer_no_lock(szone, ptr))) {
        vm_range_to_deallocate = *huge;
        *huge = szone->huge_entries[--szone->num_huge_entries]; // last entry fills that spot
     } else if ((huge = huge_entry_for_pointer_no_lock(szone, ptr))) {
        vm_range_to_deallocate = *huge;
        *huge = szone->huge_entries[--szone->num_huge_entries]; // last entry fills that spot

-       szone->num_bytes_in_huge_objects -= vm_range_to_deallocate.size;
+       szone->num_bytes_in_huge_objects -= (size_t)vm_range_to_deallocate.size;

     } else {
 #if DEBUG_MALLOC
        large_debug_print(szone);
 #endif
     } else {
 #if DEBUG_MALLOC
        large_debug_print(szone);
 #endif

-       szone_error(szone, "Pointer being freed was not allocated", ptr);
+       szone_error(szone, "pointer being freed was not allocated", ptr, NULL);

        return;
     }
     SZONE_UNLOCK(szone); // we release the lock asap
     CHECK(szone, __PRETTY_FUNCTION__);
     // we deallocate_pages, including guard pages
     if (vm_range_to_deallocate.address) {
        return;
     }
     SZONE_UNLOCK(szone); // we release the lock asap
     CHECK(szone, __PRETTY_FUNCTION__);
     // we deallocate_pages, including guard pages
     if (vm_range_to_deallocate.address) {

-//     malloc_printf("About to deallocate %p size %y\n", vm_range_to_deallocate.address, vm_range_to_deallocate.size);

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-       if (large_entry_for_pointer_no_lock(szone,
-         (void *)vm_range_to_deallocate.address)) {
-           malloc_printf("*** malloc[%d]: Invariant broken: %p still in use num_large_entries=%d\n", getpid(), vm_range_to_deallocate.address, szone->num_large_entries);
+       if (large_entry_for_pointer_no_lock(szone, (void *)vm_range_to_deallocate.address)) {
+           malloc_printf("*** invariant broken: %p still in use num_large_entries=%d\n", vm_range_to_deallocate.address, szone->num_large_entries);

            large_debug_print(szone);
            szone_sleep();
        }
 #endif
            large_debug_print(szone);
            szone_sleep();
        }
 #endif

-       deallocate_pages(szone, vm_range_to_deallocate.address, vm_range_to_deallocate.size, 0);
+       deallocate_pages(szone, (void *)vm_range_to_deallocate.address, (size_t)vm_range_to_deallocate.size, 0);

     }
 }
 
 static INLINE int
     }
 }
 
 static INLINE int

-try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size) {
+try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, size_t new_size)
+{

     vm_address_t       addr = (vm_address_t)ptr + old_size;
     vm_address_t       addr = (vm_address_t)ptr + old_size;

-    large_entry_t      *entry;
+    large_entry_t      *large_entry, saved_entry;
+    huge_entry_t       *huge_entry, huge;

     kern_return_t      err;
     kern_return_t      err;

+

 #if DEBUG_MALLOC
     if (old_size != ((old_size >> vm_page_shift) << vm_page_shift)) {
        malloc_printf("*** old_size is %d\n", old_size);
     }
 #endif
 #if DEBUG_MALLOC
     if (old_size != ((old_size >> vm_page_shift) << vm_page_shift)) {
        malloc_printf("*** old_size is %d\n", old_size);
     }
 #endif

-//    malloc_printf("=== Trying (1) to extend %p from %d to %d\n", ptr, old_size, new_size);

     SZONE_LOCK(szone);
     SZONE_LOCK(szone);

-    entry = large_entry_for_pointer_no_lock(szone, (void *)addr);
+    large_entry = large_entry_for_pointer_no_lock(szone, (void *)addr);

     SZONE_UNLOCK(szone);
     SZONE_UNLOCK(szone);

-    if (entry) {
-       return 0; // large pointer already exist in table - extension is not going to work
+    if (large_entry) {
+       return 0; // large pointer already exists in table - extension is not going to work

     }
     new_size = round_page(new_size);
     }
     new_size = round_page(new_size);

-//    malloc_printf("=== Trying (2) to extend %p from %d to %d\n", ptr, old_size, new_size);

     /*
      * Ask for allocation at a specific address, and mark as realloc
      * to request coalescing with previous realloc'ed extensions.
     /*
      * Ask for allocation at a specific address, and mark as realloc
      * to request coalescing with previous realloc'ed extensions.


@@ -2427,34 +3295,35 @@ try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, s
      * Note: this logic is predicated on the understanding that an allocated
      * block can never really shrink, so that the new size will always be 
      * larger than the old size.
      * Note: this logic is predicated on the understanding that an allocated
      * block can never really shrink, so that the new size will always be 
      * larger than the old size.

+     *
+     * Note: the use of 1 << vm_page_shift here has to do with the subdivision
+     * of the bits in the large_entry_t, and not the size of a page (directly).

      */
     if ((new_size >> vm_page_shift) < (1 << vm_page_shift)) {
        /* extend existing large entry */
      */
     if ((new_size >> vm_page_shift) < (1 << vm_page_shift)) {
        /* extend existing large entry */

-       entry = large_entry_for_pointer_no_lock(szone, ptr);
-       if (!entry) {
-           szone_error(szone, "large entry reallocated is not properly in table", ptr);
+       large_entry = large_entry_for_pointer_no_lock(szone, ptr);
+       if (!large_entry) {
+           szone_error(szone, "large entry reallocated is not properly in table", ptr, NULL);

            /* XXX will cause fault on next reference to entry */
        }
            /* XXX will cause fault on next reference to entry */
        }

-       entry->address_and_num_pages = (vm_address_t)ptr | (new_size >> vm_page_shift);
+       large_entry->address_and_num_pages = (uintptr_t)ptr | (new_size >> vm_page_shift);

        szone->num_bytes_in_large_objects += new_size - old_size;
     } else if ((old_size >> vm_page_shift) >= (1 << vm_page_shift)) {
        /* extend existing huge entry */
        szone->num_bytes_in_large_objects += new_size - old_size;
     } else if ((old_size >> vm_page_shift) >= (1 << vm_page_shift)) {
        /* extend existing huge entry */

-       huge_entry_t    *huge_entry = huge_entry_for_pointer_no_lock(szone, ptr);
+       huge_entry = huge_entry_for_pointer_no_lock(szone, ptr);

        if (!huge_entry) {
        if (!huge_entry) {

-           szone_error(szone, "huge entry reallocated is not properly in table", ptr);
+           szone_error(szone, "huge entry reallocated is not properly in table", ptr, NULL);

            /* XXX will cause fault on next reference to huge_entry */
        }
        huge_entry->size = new_size;
        szone->num_bytes_in_huge_objects += new_size - old_size;
     } else {
        /* need to convert large entry to huge entry */
            /* XXX will cause fault on next reference to huge_entry */
        }
        huge_entry->size = new_size;
        szone->num_bytes_in_huge_objects += new_size - old_size;
     } else {
        /* need to convert large entry to huge entry */

-       huge_entry_t huge;

 
        /* release large entry, note we still have the VM allocation */
 
        /* release large entry, note we still have the VM allocation */

-       entry = large_entry_for_pointer_no_lock(szone, ptr);
-       large_entry_t   saved_entry = *entry; // in case we need to put it back
-       large_free_no_lock(szone, entry);
-       szone->num_bytes_in_large_objects -= old_size;
+       large_entry = large_entry_for_pointer_no_lock(szone, ptr);
+       saved_entry = *large_entry; // in case we need to put it back
+       large_free_no_lock(szone, large_entry);

 
        /* and get a huge entry */
        huge.address = (vm_address_t)ptr;
 
        /* and get a huge entry */
        huge.address = (vm_address_t)ptr;


@@ -2468,7 +3337,6 @@ try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, s
        // we leak memory (the extra space appended) but data structures are correct
        large_entry_insert_no_lock(szone, saved_entry); // this will reinsert the large entry
     }
        // we leak memory (the extra space appended) but data structures are correct
        large_entry_insert_no_lock(szone, saved_entry); // this will reinsert the large entry
     }

-//    malloc_printf("=== Successfully reallocated at end of %p from %d to %d\n", ptr, old_size, new_size);

     SZONE_UNLOCK(szone); // we release the lock asap
     return 1;
 }
     SZONE_UNLOCK(szone); // we release the lock asap
     return 1;
 }


@@ -2476,124 +3344,174 @@ try_realloc_large_or_huge_in_place(szone_t *szone, void *ptr, size_t old_size, s
 /********************* Zone call backs ************************/
 
 static void
 /********************* Zone call backs ************************/
 
 static void

-szone_free(szone_t *szone, void *ptr) {
-    // malloc_printf("szone_free(%p)\n", ptr);
+szone_free(szone_t *szone, void *ptr)
+{
+    region_t   *tiny_region;
+    region_t   *small_region;
+

 #if DEBUG_MALLOC
 #if DEBUG_MALLOC

-    if (LOG(szone, ptr)) malloc_printf("In szone_free with %p\n", ptr);
+    if (LOG(szone, ptr))
+       malloc_printf("in szone_free with %p\n", ptr);

 #endif
 #endif

-    if (!ptr) return;
-    if ((vm_address_t)ptr & (TINY_QUANTUM - 1)) {
-       szone_error(szone, "Non-aligned pointer being freed", ptr);
+    if (!ptr)
+       return;
+    /*
+     * Try to free to a tiny region.
+     */
+    if ((uintptr_t)ptr & (TINY_QUANTUM - 1)) {
+       szone_error(szone, "Non-aligned pointer being freed", ptr, NULL);

        return;
     }
        return;
     }

-    // try a tiny pointer
-    tiny_region_t      *tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
-    if (tiny_region) {
+    if ((tiny_region = tiny_region_for_ptr_no_lock(szone, ptr)) != NULL) {
+       if (TINY_INDEX_FOR_PTR(ptr) >= NUM_TINY_BLOCKS) {
+           szone_error(szone, "Pointer to metadata being freed", ptr, NULL);
+           return;
+       }

        free_tiny(szone, ptr, tiny_region);
        return;
     }
        free_tiny(szone, ptr, tiny_region);
        return;
     }

-    if ((vm_address_t)ptr & (SMALL_QUANTUM - 1)) {
-       szone_error(szone, "Non-aligned pointer being freed (2)", ptr);
+
+    /*
+     * Try to free to a small region.
+     */
+    if ((uintptr_t)ptr & (SMALL_QUANTUM - 1)) {
+       szone_error(szone, "Non-aligned pointer being freed (2)", ptr, NULL);

        return;
     }
        return;
     }

-    // try a small pointer
-    small_region_t             *small_region = small_region_for_ptr_no_lock(szone, ptr);
-    if (small_region) {
+    if ((small_region = small_region_for_ptr_no_lock(szone, ptr)) != NULL) {
+       if (SMALL_META_INDEX_FOR_PTR(ptr) >= NUM_SMALL_BLOCKS) {
+           szone_error(szone, "Pointer to metadata being freed (2)", ptr, NULL);
+           return;
+       }

        free_small(szone, ptr, small_region);
        return;
     }
        free_small(szone, ptr, small_region);
        return;
     }

-    if (((unsigned)ptr) & ((1 << vm_page_shift) - 1)) {
-       szone_error(szone, "Non-page-aligned, non-allocated pointer being freed", ptr);
+
+    /* check that it's a legal large/huge allocation */
+    if ((uintptr_t)ptr & (vm_page_size - 1)) {
+       szone_error(szone, "non-page-aligned, non-allocated pointer being freed", ptr, NULL);

        return;
     }
     free_large_or_huge(szone, ptr);
 }
 
 static INLINE void *
        return;
     }
     free_large_or_huge(szone, ptr);
 }
 
 static INLINE void *

-szone_malloc_should_clear(szone_t *szone, size_t size, boolean_t cleared_requested) {
+szone_malloc_should_clear(szone_t *szone, size_t size, boolean_t cleared_requested)
+{

     void       *ptr;
     void       *ptr;

+    msize_t    msize;
+

     if (size <= 31*TINY_QUANTUM) {
        // think tiny
     if (size <= 31*TINY_QUANTUM) {
        // think tiny

-       msize_t         msize = (size + TINY_QUANTUM - 1) >> SHIFT_TINY_QUANTUM;
-       if (! msize) msize = 1;
+       msize = TINY_MSIZE_FOR_BYTES(size + TINY_QUANTUM - 1);
+       if (!msize)
+           msize = 1;

        ptr = tiny_malloc_should_clear(szone, msize, cleared_requested);
     } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (size < LARGE_THRESHOLD)) {
        // think small
        ptr = tiny_malloc_should_clear(szone, msize, cleared_requested);
     } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (size < LARGE_THRESHOLD)) {
        // think small

-       msize_t         msize = (size + SMALL_QUANTUM - 1) >> SHIFT_SMALL_QUANTUM;
+       msize = SMALL_MSIZE_FOR_BYTES(size + SMALL_QUANTUM - 1);

        if (! msize) msize = 1;
        ptr = small_malloc_should_clear(szone, msize, cleared_requested);
     } else {
        if (! msize) msize = 1;
        ptr = small_malloc_should_clear(szone, msize, cleared_requested);
     } else {

-       unsigned                num_pages;
-       num_pages = round_page(size) >> vm_page_shift;
+       // large or huge
+       size_t num_pages = round_page(size) >> vm_page_shift;
+       if (num_pages == 0)     /* Overflowed */
+               ptr = 0;
+       else

        ptr = large_and_huge_malloc(szone, num_pages);
     }
 #if DEBUG_MALLOC
        ptr = large_and_huge_malloc(szone, num_pages);
     }
 #if DEBUG_MALLOC

-    if (LOG(szone, ptr)) malloc_printf("szone_malloc returned %p\n", ptr);
+    if (LOG(szone, ptr))
+       malloc_printf("szone_malloc returned %p\n", ptr);

 #endif
 #endif

+    /*
+     * If requested, scribble on allocated memory.
+     */
+    if ((szone->debug_flags & SCALABLE_MALLOC_DO_SCRIBBLE) && ptr && !cleared_requested && size)
+       memset(ptr, 0xaa, size);
+

     return ptr;
 }
 
 static void *
 szone_malloc(szone_t *szone, size_t size) {
     return ptr;
 }
 
 static void *
 szone_malloc(szone_t *szone, size_t size) {

-    // malloc_printf("szone_malloc(%d)\n", size);
-    void *ptr = szone_malloc_should_clear(szone, size, 0);
-    // malloc_printf("szone_malloc(%d) -> %p %d\n", size, ptr, malloc_size(ptr));
-    return ptr;
+    return szone_malloc_should_clear(szone, size, 0);

 }
 
 static void *
 }
 
 static void *

-szone_calloc(szone_t *szone, size_t num_items, size_t size) {
-    // malloc_printf("szone_calloc(%d,%d)\n", num_items, size);
-    void       *ptr = szone_malloc_should_clear(szone, num_items * size, 1);
-    // malloc_printf("szone_calloc(%d,%d) -> %p\n", num_items, size, ptr);
-    return ptr;
+szone_calloc(szone_t *szone, size_t num_items, size_t size)
+{
+    size_t total_bytes = num_items * size;
+    if ((num_items > 1) && (size != 0) && ((total_bytes / size) != num_items))
+        return NULL;
+    return szone_malloc_should_clear(szone, total_bytes, 1);

 }
 
 static void *
 }
 
 static void *

-szone_valloc(szone_t *szone, size_t size) {
+szone_valloc(szone_t *szone, size_t size)
+{

     void       *ptr;
     void       *ptr;

-    unsigned   num_pages;
+    size_t num_pages;
+    

     num_pages = round_page(size) >> vm_page_shift;
     ptr = large_and_huge_malloc(szone, num_pages);
 #if DEBUG_MALLOC
     num_pages = round_page(size) >> vm_page_shift;
     ptr = large_and_huge_malloc(szone, num_pages);
 #if DEBUG_MALLOC

-    if (LOG(szone, ptr)) malloc_printf("szone_valloc returned %p\n", ptr);
+    if (LOG(szone, ptr))
+       malloc_printf("szone_valloc returned %p\n", ptr);

 #endif
     return ptr;
 }
 
 static size_t
 #endif
     return ptr;
 }
 
 static size_t

-szone_size(szone_t *szone, const void *ptr) {
+szone_size(szone_t *szone, const void *ptr)
+{

     size_t             size = 0;
     size_t             size = 0;

+    boolean_t          is_free;
+    msize_t            msize, msize_and_free;

     large_entry_t      *entry;
     huge_entry_t       *huge;
     large_entry_t      *entry;
     huge_entry_t       *huge;

-    // malloc_printf("szone_size(%p)\n", ptr);
-    if (!ptr) return 0;
+    
+    if (!ptr)
+       return 0;

 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {

-       malloc_printf("In szone_size for %p (szone=%p)\n", ptr, szone);
-    }
-#endif
-    if ((vm_address_t)ptr & (TINY_QUANTUM - 1)) return 0;
-    // Try tiny
-    tiny_region_t      *tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
-    if (tiny_region) {
-       // this is indeed a valid pointer
-       boolean_t       is_free;
-       msize_t         msize = get_tiny_meta_header(ptr, &is_free);
-       return (is_free) ? 0 : msize << SHIFT_TINY_QUANTUM;
-    }
-    if ((vm_address_t)ptr & (SMALL_QUANTUM - 1)) return 0;
-    // Try a small
-    small_region_t             *small_region = small_region_for_ptr_no_lock(szone, ptr);
-    if (small_region) {
-       // this is indeed a valid pointer
-       msize_t         msize_and_free = small_meta_header(ptr)[0];
-       return (msize_and_free & SMALL_IS_FREE) ? 0 : msize_and_free << SHIFT_SMALL_QUANTUM;
-    }
-    if (((unsigned)ptr) & ((1 << vm_page_shift) - 1)) {
-       // malloc_printf("Object %p not found in szone_size\n", ptr);
+       malloc_printf("in szone_size for %p (szone=%p)\n", ptr, szone);
+    }
+#endif
+
+    /*
+     * Look for it in a tiny region.
+     */
+    if ((uintptr_t)ptr & (TINY_QUANTUM - 1))
+       return 0;
+    if (tiny_region_for_ptr_no_lock(szone, ptr)) {
+       if (TINY_INDEX_FOR_PTR(ptr) >= NUM_TINY_BLOCKS)
+           return 0;
+       msize = get_tiny_meta_header(ptr, &is_free);
+       return (is_free) ? 0 : TINY_BYTES_FOR_MSIZE(msize);
+    }
+    
+    /*
+     * Look for it in a small region.
+     */
+    if ((uintptr_t)ptr & (SMALL_QUANTUM - 1))

        return 0;
        return 0;

+    if (small_region_for_ptr_no_lock(szone, ptr)) {
+       if (SMALL_META_INDEX_FOR_PTR(ptr) >= NUM_SMALL_BLOCKS)
+           return 0;
+       msize_and_free = *SMALL_METADATA_FOR_PTR(ptr);
+       return (msize_and_free & SMALL_IS_FREE) ? 0 : SMALL_BYTES_FOR_MSIZE(msize_and_free);

     }
     }

+
+    /*
+     * If not page-aligned, it cannot have come from a large or huge allocation.
+     */
+    if ((uintptr_t)ptr & (vm_page_size - 1))
+       return(0);
+
+    /*
+     * Look for it in a large or huge entry.
+     */

     SZONE_LOCK(szone);
     entry = large_entry_for_pointer_no_lock(szone, ptr);
     if (entry) {
     SZONE_LOCK(szone);
     entry = large_entry_for_pointer_no_lock(szone, ptr);
     if (entry) {


@@ -2602,7 +3520,6 @@ szone_size(szone_t *szone, const void *ptr) {
        size = huge->size;
     }
     SZONE_UNLOCK(szone); 
        size = huge->size;
     }
     SZONE_UNLOCK(szone); 

-    // malloc_printf("szone_size for large/huge %p returned %d\n", ptr, (unsigned)size);

 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
        malloc_printf("szone_size for %p returned %d\n", ptr, (unsigned)size);
 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
        malloc_printf("szone_size for %p returned %d\n", ptr, (unsigned)size);


@@ -2612,203 +3529,238 @@ szone_size(szone_t *szone, const void *ptr) {
 }
 
 static void *
 }
 
 static void *

-szone_realloc(szone_t *szone, void *ptr, size_t new_size) {
-    size_t             old_size = 0;
+szone_realloc(szone_t *szone, void *ptr, size_t new_size)
+{
+    size_t             old_size;

     void               *new_ptr;
     void               *new_ptr;

-    // malloc_printf("szone_realloc(%p,%d)\n", ptr, new_size);
+    

 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {

-       malloc_printf("In szone_realloc for %p, %d\n", ptr, (unsigned)new_size);
+       malloc_printf("in szone_realloc for %p, %d\n", ptr, (unsigned)new_size);

     }
 #endif
     if (!ptr) {
        ptr = szone_malloc(szone, new_size);
     }
 #endif
     if (!ptr) {
        ptr = szone_malloc(szone, new_size);

-       // malloc_printf("szone_realloc(%p,%d) -> %p\n", ptr, new_size, ptr);

        return ptr;
     }
     old_size = szone_size(szone, ptr);
     if (!old_size) {
        return ptr;
     }
     old_size = szone_size(szone, ptr);
     if (!old_size) {

-       szone_error(szone, "Pointer being reallocated was not allocated", ptr);
+       szone_error(szone, "pointer being reallocated was not allocated", ptr, NULL);

        return NULL;
     }
     /* we never shrink an allocation */
        return NULL;
     }
     /* we never shrink an allocation */

-    if (old_size >= new_size) return ptr;
+    if (old_size >= new_size)
+       return ptr;
+
+    /*
+     * If the old and new sizes both suit the tiny allocator, try to reallocate in-place.
+     */

     if ((new_size + TINY_QUANTUM - 1) <= 31 * TINY_QUANTUM) {
     if ((new_size + TINY_QUANTUM - 1) <= 31 * TINY_QUANTUM) {

-       // We now try to realloc in place

        if (try_realloc_tiny_in_place(szone, ptr, old_size, new_size)) {
        if (try_realloc_tiny_in_place(szone, ptr, old_size, new_size)) {

-           // malloc_printf("szone_realloc(%p,%d) -> %p\n", ptr, new_size, ptr);

            return ptr;
        }
            return ptr;
        }

-    } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && ((new_size + SMALL_QUANTUM - 1) < LARGE_THRESHOLD) && (old_size > 31 * TINY_QUANTUM)) {
-       // We now try to realloc in place
+
+       /*
+        * If the old and new sizes both suit the small allocator, and we're not protecting the
+        * small allocations, try to reallocate in-place.
+        */
+    } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) &&
+      ((new_size + SMALL_QUANTUM - 1) < LARGE_THRESHOLD) &&
+      (old_size > 31 * TINY_QUANTUM)) {

        if (try_realloc_small_in_place(szone, ptr, old_size, new_size)) {
        if (try_realloc_small_in_place(szone, ptr, old_size, new_size)) {

-           // malloc_printf("szone_realloc(%p,%d) small in place -> %p\n", ptr, new_size, ptr);

            return ptr;
        }
            return ptr;
        }

+
+       /*
+        * If the allocation's a large or huge allocation, try to reallocate in-place there.
+        */

     } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (old_size > LARGE_THRESHOLD)) {
        if (try_realloc_large_or_huge_in_place(szone, ptr, old_size, new_size)) {
            return ptr;
        }
     }
     } else if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (old_size > LARGE_THRESHOLD)) {
        if (try_realloc_large_or_huge_in_place(szone, ptr, old_size, new_size)) {
            return ptr;
        }
     }

+
+    /*
+     * Can't reallocate in place for whatever reason; allocate a new buffer and copy.
+     */

     new_ptr = szone_malloc(szone, new_size);
     new_ptr = szone_malloc(szone, new_size);

-    if (new_ptr == NULL) return NULL;
-    if ((old_size > VM_COPY_THRESHOLD) && (new_size > VM_COPY_THRESHOLD)) {
-       // we know everything is page-aligned try vm_copy
-       kern_return_t   err = 0;
-       err = vm_copy(mach_task_self(), (vm_address_t)ptr, old_size, (vm_address_t)new_ptr);
-       if (err) {
-           szone_error(szone, "Can't vm_copy region", ptr);
-       }
-    } else {
+    if (new_ptr == NULL)
+       return NULL;
+
+    /*
+     * If the allocation's large enough, try to copy using VM.  If that fails, or
+     * if it's too small, just copy by hand.
+     */
+    if ((old_size < VM_COPY_THRESHOLD) ||
+      vm_copy(mach_task_self(), (vm_address_t)ptr, old_size, (vm_address_t)new_ptr))

        memcpy(new_ptr, ptr, old_size);
        memcpy(new_ptr, ptr, old_size);

-    }

     szone_free(szone, ptr);
     szone_free(szone, ptr);

+    

 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
        malloc_printf("szone_realloc returned %p for %d\n", new_ptr, (unsigned)new_size);
     }
 #endif
 #if DEBUG_MALLOC
     if (LOG(szone, ptr)) {
        malloc_printf("szone_realloc returned %p for %d\n", new_ptr, (unsigned)new_size);
     }
 #endif

-    // malloc_printf("szone_realloc(%p,%d) -> %p\n", ptr, new_size, new_ptr);

     return new_ptr;
 }
 
     return new_ptr;
 }
 

-unsigned
-szone_batch_malloc(szone_t *szone, size_t size, void **results, unsigned count) {
-    // given a size, returns pointers capable of holding that size
-    // returns the number of pointers allocated
-    // may return 0 - this function will do best attempts, but just that
-    // malloc_printf("In szone_batch_malloc(%d, %d)\n", size, count);
-    if (size > 31*TINY_QUANTUM) return 0; // only bother implementing this for tiny
-    msize_t            msize = (size + TINY_QUANTUM - 1) >> SHIFT_TINY_QUANTUM;
-    if (! msize) msize = 1;
-    size_t             chunk_size = msize << SHIFT_TINY_QUANTUM;
-    unsigned   found = 0;
+// given a size, returns the number of pointers allocated capable of holding
+// that size, up to the limit specified by the 'count' argument.  These pointers
+// are stored in the 'results' array, which must be allocated by the caller.
+// may return zero, since this function is only a best attempt at allocating
+// the pointers.  clients should be prepared to call malloc for any additional
+// blocks they need.
+static unsigned
+szone_batch_malloc(szone_t *szone, size_t size, void **results, unsigned count)
+{
+    msize_t     msize = TINY_MSIZE_FOR_BYTES(size + TINY_QUANTUM - 1);
+    unsigned    found = 0;
+
+    // only bother implementing this for tiny
+    if (size > 31*TINY_QUANTUM)
+        return 0;
+    // make sure to return objects at least one quantum in size
+    if (!msize)
+        msize = 1;
+

     CHECK(szone, __PRETTY_FUNCTION__);
     CHECK(szone, __PRETTY_FUNCTION__);

-    SZONE_LOCK(szone); // might as well lock right here to avoid concurrency issues
-    free_list_t        **free_list = szone->tiny_free_list + msize - 1;
-    free_list_t        *ptr = *free_list;
-    while (found < count) {
-       if (!ptr) break;
-       *results++ = ptr; found++;
-       set_tiny_meta_header_in_use(ptr, msize);
-       ptr = ((free_list_t *)ptr)->next;
+
+       // We must lock the zone now, since tiny_malloc_from_free_list assumes that 
+       // the caller has done so.
+    SZONE_LOCK(szone);
+
+       // with the zone locked, allocate objects from the free list until all
+       // sufficiently large objects have been exhausted, or we have met our quota
+       // of objects to allocate.
+       while (found < count) {
+        void *ptr = tiny_malloc_from_free_list(szone, msize);
+        if (!ptr)
+            break;
+
+        *results++ = ptr;
+        found++;

     }
     }

-    if (ptr) {
-       ((free_list_t *)ptr)->previous = NULL;
-       free_list_set_checksum(szone, (free_list_t *)ptr);
-    }
-    *free_list = (void *)ptr;
-    // Note that we could allocate from the free lists for larger msize
-    // But that may un-necessarily fragment - so we might as well let the client do that
-    // We could also allocate from szone->tiny_bytes_free_at_end
-    // But that means we'll "eat-up" the untouched area faster, increasing the working set
-    // So we just return what we have and just that
-    szone->num_tiny_objects += found;
-    szone->num_bytes_in_tiny_objects += chunk_size * found;

     SZONE_UNLOCK(szone);
     SZONE_UNLOCK(szone);

-    // malloc_printf("In szone_batch_malloc(%d, %d) -> %d\n", size, count, found);

     return found;
 }
 
     return found;
 }
 

-void
-szone_batch_free(szone_t *szone, void **to_be_freed, unsigned count) {
+static void
+szone_batch_free(szone_t *szone, void **to_be_freed, unsigned count)
+{
+    unsigned           cc = 0;
+    void               *ptr;
+    region_t   *tiny_region;
+    boolean_t          is_free;
+    msize_t            msize;
+

     // frees all the pointers in to_be_freed
     // note that to_be_freed may be overwritten during the process
     // frees all the pointers in to_be_freed
     // note that to_be_freed may be overwritten during the process

-    if (!count) return;
-    // malloc_printf("Freeing %d items\n", count);
-    unsigned   cc = 0;
+    if (!count)
+       return;

     CHECK(szone, __PRETTY_FUNCTION__);
     SZONE_LOCK(szone);
     while (cc < count) {
     CHECK(szone, __PRETTY_FUNCTION__);
     SZONE_LOCK(szone);
     while (cc < count) {

-       void    *ptr = to_be_freed[cc];
-       tiny_region_t   *tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
-       if (tiny_region) {
-           // this is a tiny pointer
-           boolean_t   is_free;
-           msize_t     msize = get_tiny_meta_header(ptr, &is_free);
-           if (is_free) break; // a double free; let the standard free deal with it
-           tiny_free_no_lock(szone, tiny_region, ptr, msize);
-           to_be_freed[cc] = NULL;
+       ptr = to_be_freed[cc];
+       if (ptr) {
+           /* XXX this really slows us down */
+           tiny_region = tiny_region_for_ptr_no_lock(szone, ptr);
+           if (tiny_region) {
+               // this is a tiny pointer
+               if (TINY_INDEX_FOR_PTR(ptr) >= NUM_TINY_BLOCKS)
+                   break; // pointer to metadata; let the standard free deal with it
+               msize = get_tiny_meta_header(ptr, &is_free);
+               if (is_free)
+                   break; // a double free; let the standard free deal with it
+               tiny_free_no_lock(szone, tiny_region, ptr, msize);
+               to_be_freed[cc] = NULL;
+           }

        }
        cc++;
     }
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
     while (count--) {
        }
        cc++;
     }
     SZONE_UNLOCK(szone);
     CHECK(szone, __PRETTY_FUNCTION__);
     while (count--) {

-       void    *ptr = to_be_freed[count];
-       // malloc_printf("Freeing item at %d: %p\n", count, ptr);
-       if (ptr) szone_free(szone, ptr);
+       ptr = to_be_freed[count];
+       if (ptr)
+           szone_free(szone, ptr);

     }
 }
 
 static void
     }
 }
 
 static void

-szone_destroy(szone_t *szone) {
-    unsigned   index;
-    small_region_t     pended_region = 0;
+szone_destroy(szone_t *szone)
+{
+    size_t             index;
+    large_entry_t      *large;
+    vm_range_t         range_to_deallocate;
+    huge_entry_t       *huge;
+
+    /* destroy large entries */

     index = szone->num_large_entries;
     while (index--) {
     index = szone->num_large_entries;
     while (index--) {

-       large_entry_t   *entry = szone->large_entries + index;
-       if (!LARGE_ENTRY_IS_EMPTY(*entry)) {
-           large_entry_t       range;
-           range = *entry;
+       large = szone->large_entries + index;
+       if (!LARGE_ENTRY_IS_EMPTY(*large)) {

            // we deallocate_pages, including guard pages
            // we deallocate_pages, including guard pages

-           deallocate_pages(szone, LARGE_ENTRY_ADDRESS(range), LARGE_ENTRY_SIZE(range), szone->debug_flags);
+           deallocate_pages(szone, (void *)LARGE_ENTRY_ADDRESS(*large), LARGE_ENTRY_SIZE(*large), szone->debug_flags);

        }
     }
     if (szone->num_large_entries * sizeof(large_entry_t) >= LARGE_THRESHOLD) {
        }
     }
     if (szone->num_large_entries * sizeof(large_entry_t) >= LARGE_THRESHOLD) {

-       vm_range_t      range_to_deallocate;
-       large_entries_free_no_lock(szone, szone->large_entries, szone->num_large_entries, &range_to_deallocate); // we do not free in the small chunk case
-       if (range_to_deallocate.size) deallocate_pages(szone, range_to_deallocate.address, range_to_deallocate.size, 0);
-
+       // we do not free in the small chunk case
+       large_entries_free_no_lock(szone, szone->large_entries, szone->num_large_entries, &range_to_deallocate);
+       if (range_to_deallocate.size)
+           deallocate_pages(szone, (void *)range_to_deallocate.address, (size_t)range_to_deallocate.size, 0);

     }
     }

+
+    /* destroy huge entries */

     index = szone->num_huge_entries;
     while (index--) {
     index = szone->num_huge_entries;
     while (index--) {

-       huge_entry_t    *huge = szone->huge_entries + index;
-       deallocate_pages(szone, huge->address, huge->size, szone->debug_flags);
-    }
-    // the tiny regions
-    index = szone->num_tiny_regions;
-    while (index--) {
-        tiny_region_t  tiny_region = szone->tiny_regions[index];
-       vm_size_t       size_allocated = ((TINY_REGION_SIZE + (1 << vm_page_shift) - 1) >> vm_page_shift) << vm_page_shift;
-       deallocate_pages(szone, TINY_REGION_ADDRESS(tiny_region), size_allocated, 0);
-    }
-    // and now we free regions, with regions[0] as the last one (the final harakiri)
-    index = szone->num_small_regions;
-    while (index--) {
-        small_region_t region = szone->small_regions[index];
-       if (index > 0
-           && (void *)szone->small_regions >= (void *)(SMALL_REGION_ADDRESS(region))
-           && (void *)szone->small_regions < (void *)(SMALL_REGION_END(region))) {
-               // Pend deallocation of this region, since the region
-               // bookkeeping array is in it.
-               pended_region = region;
-       } else {
-               deallocate_pages(szone, SMALL_REGION_ADDRESS(region), SMALL_REGION_SIZE, 0);
-       }
-    }
-    if (pended_region) {
-        deallocate_pages(szone, SMALL_REGION_ADDRESS(pended_region), SMALL_REGION_SIZE, 0);
+       huge = szone->huge_entries + index;
+       deallocate_pages(szone, (void *)huge->address, huge->size, szone->debug_flags);

     }
     }

+    
+    /* destroy tiny regions */
+    for (index = 0; index < szone->num_tiny_regions_allocated; ++index)
+      if (szone->tiny_regions[index])
+        deallocate_pages(szone, szone->tiny_regions[index], TINY_REGION_SIZE, 0);
+
+    /* destroy small regions */
+    for (index = 0; index < szone->num_small_regions_allocated; ++index)
+      if (szone->small_regions[index])
+        deallocate_pages(szone, szone->small_regions[index], SMALL_REGION_SIZE, 0);
+
+    /* destroy region hash rings, if any */
+    if (szone->tiny_regions != szone->initial_tiny_regions) {
+      size_t size = round_page(szone->num_tiny_regions_allocated * sizeof(region_t));
+      deallocate_pages(szone, szone->tiny_regions, size, 0);
+    }
+    if (szone->small_regions != szone->initial_small_regions) {
+      size_t size = round_page(szone->num_small_regions_allocated * sizeof(region_t));
+      deallocate_pages(szone, szone->small_regions, size, 0);
+    }
+    /* Now destroy the separate szone region */
+    deallocate_pages(szone, (void *)szone, SZONE_PAGED_SIZE, SCALABLE_MALLOC_ADD_GUARD_PAGES);

 }
 
 static size_t
 }
 
 static size_t

-szone_good_size(szone_t *szone, size_t size) {
+szone_good_size(szone_t *szone, size_t size)
+{
+    msize_t    msize;
+    unsigned   num_pages;
+    

     if (size <= 31 * TINY_QUANTUM) {
        // think tiny
     if (size <= 31 * TINY_QUANTUM) {
        // think tiny

-       msize_t msize = (size + TINY_QUANTUM - 1) >> SHIFT_TINY_QUANTUM;
+       msize = TINY_MSIZE_FOR_BYTES(size + TINY_QUANTUM - 1);

        if (! msize) msize = 1;
        if (! msize) msize = 1;

-       return msize << SHIFT_TINY_QUANTUM;
+       return TINY_BYTES_FOR_MSIZE(msize);

     }
     if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (size < LARGE_THRESHOLD)) {
        // think small
     }
     if (!((szone->debug_flags & SCALABLE_MALLOC_ADD_GUARD_PAGES) && PROTECT_SMALL) && (size < LARGE_THRESHOLD)) {
        // think small

-       msize_t msize = (size + SMALL_QUANTUM - 1) >> SHIFT_SMALL_QUANTUM;
+       msize = SMALL_MSIZE_FOR_BYTES(size + SMALL_QUANTUM - 1);

        if (! msize) msize = 1;
        if (! msize) msize = 1;

-       return msize << SHIFT_SMALL_QUANTUM;
+       return SMALL_BYTES_FOR_MSIZE(msize);

     } else {
     } else {

-       unsigned                num_pages;

        num_pages = round_page(size) >> vm_page_shift;
        num_pages = round_page(size) >> vm_page_shift;

-       if (!num_pages) num_pages = 1; // minimal allocation size for this
+       if (!num_pages)
+           num_pages = 1; // minimal allocation size for this

        return num_pages << vm_page_shift;
     }
 }
        return num_pages << vm_page_shift;
     }
 }


@@ -2818,93 +3770,92 @@ unsigned szone_check_start = 0;
 unsigned szone_check_modulo = 1;
 
 static boolean_t
 unsigned szone_check_modulo = 1;
 
 static boolean_t

-szone_check_all(szone_t *szone, const char *function) {
-    unsigned   index = 0;
+szone_check_all(szone_t *szone, const char *function)
+{
+    size_t index;
+

     SZONE_LOCK(szone);
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);
     SZONE_LOCK(szone);
     CHECK_LOCKED(szone, __PRETTY_FUNCTION__);

-    while (index < szone->num_tiny_regions) {
-       tiny_region_t   *region = szone->tiny_regions + index++;
-       if (! szone_check_tiny_region(szone, region)) {
-           SZONE_UNLOCK(szone);
-           szone->debug_flags &= ~ CHECK_REGIONS;
-           malloc_printf("*** malloc[%d]: Tiny region %d incorrect szone_check_all(%s) counter=%d\n", getpid(), index-1, function, szone_check_counter);
-           szone_error(szone, "Check: tiny region incorrect", NULL);
-           return 0;
-       }
-    }
-
-    index = 0;
-    while (index < NUM_TINY_SLOTS) {
-       if (! tiny_free_list_check(szone, index)) {
-           SZONE_UNLOCK(szone);
-           szone->debug_flags &= ~ CHECK_REGIONS;
-           malloc_printf("*** malloc[%d]: Tiny free list incorrect (slot=%d) szone_check_all(%s) counter=%d\n", getpid(), index, function, szone_check_counter);
-           szone_error(szone, "Check: tiny free list incorrect", NULL);
-           return 0;
-       }
-       index++;
-    }

 
 

-    index = 0; while (index < szone->num_small_regions) {
-       small_region_t  *region = szone->small_regions + index++;
-       if (! szone_check_small_region(szone, region)) {
-           SZONE_UNLOCK(szone);
-           szone->debug_flags &= ~ CHECK_REGIONS;
-           malloc_printf("*** malloc[%d]: Small region %d incorrect szone_check_all(%s) counter=%d\n", getpid(), index-1, function, szone_check_counter);
-           szone_error(szone, "Check: small region incorrect", NULL);
-           return 0;
-       }
-    }
-    index = 0;
-    while (index < NUM_SMALL_SLOTS) {
-       if (! small_free_list_check(szone, index)) {
-           SZONE_UNLOCK(szone);
-           szone->debug_flags &= ~ CHECK_REGIONS;
-           malloc_printf("*** malloc[%d]: Small free list incorrect (grain=%d) szone_check_all(%s) counter=%d\n", getpid(), index, function, szone_check_counter);
-           szone_error(szone, "Check: small free list incorrect", NULL);
-           return 0;
-       }
-       index++;
+    /* check tiny regions - chould check region count */
+    for (index = 0; index < szone->num_tiny_regions_allocated; ++index) {
+      region_t tiny = szone->tiny_regions[index];
+      if (tiny && !tiny_check_region(szone, tiny)) {
+        SZONE_UNLOCK(szone);
+        szone->debug_flags &= ~ CHECK_REGIONS;
+        szone_error(szone, "check: tiny region incorrect", NULL,
+                    "*** tiny region %d incorrect szone_check_all(%s) counter=%d\n",
+                    index, function, szone_check_counter);
+        return 0;
+      }
+    }
+    /* check tiny free lists */
+    for (index = 0; index < NUM_TINY_SLOTS; ++index) {
+      if (!tiny_free_list_check(szone, index)) {
+        SZONE_UNLOCK(szone);
+        szone->debug_flags &= ~ CHECK_REGIONS;
+        szone_error(szone, "check: tiny free list incorrect", NULL,
+                    "*** tiny free list incorrect (slot=%d) szone_check_all(%s) counter=%d\n",
+                    index, function, szone_check_counter);
+        return 0;
+      }
+    }
+    /* check small regions - could check region count */
+    for (index = 0; index < szone->num_small_regions_allocated; ++index) {
+      region_t small = szone->small_regions[index];
+      if (small && !szone_check_small_region(szone, small)) {
+        SZONE_UNLOCK(szone);
+        szone->debug_flags &= ~ CHECK_REGIONS;
+        szone_error(szone, "check: small region incorrect", NULL,
+                    "*** small region %d incorrect szone_check_all(%s) counter=%d\n",
+                    index, function, szone_check_counter);
+        return 0;
+      }
+    }
+    /* check small free lists */
+    for (index = 0; index < NUM_SMALL_SLOTS; ++index) {
+      if (!small_free_list_check(szone, index)) {
+        SZONE_UNLOCK(szone);
+        szone->debug_flags &= ~ CHECK_REGIONS;
+        szone_error(szone, "check: small free list incorrect", NULL,
+                    "*** small free list incorrect (grain=%d) szone_check_all(%s) counter=%d\n",
+                    index, function, szone_check_counter);
+        return 0;
+      }

     }
     SZONE_UNLOCK(szone);
     }
     SZONE_UNLOCK(szone);

-    // szone_print(szone, 1);

     return 1;
 }
 
 static boolean_t
     return 1;
 }
 
 static boolean_t

-szone_check(szone_t *szone) {
-    if (! (++szone_check_counter % 10000)) {
-       malloc_printf("At szone_check counter=%d\n", szone_check_counter);
-    }
-    if (szone_check_counter < szone_check_start) return 1;
-    if (szone_check_counter % szone_check_modulo) return 1;
+szone_check(szone_t *szone)
+{
+    if ((++szone_check_counter % 10000) == 0)
+       _malloc_printf(ASL_LEVEL_NOTICE, "at szone_check counter=%d\n", szone_check_counter);
+    if (szone_check_counter < szone_check_start)
+       return 1;
+    if (szone_check_counter % szone_check_modulo)
+       return 1;

     return szone_check_all(szone, "");
 }
 
 static kern_return_t
     return szone_check_all(szone, "");
 }
 
 static kern_return_t

-szone_ptr_in_use_enumerator(task_t task, void *context,
-  unsigned type_mask, vm_address_t zone_address, memory_reader_t reader,
-  vm_range_recorder_t recorder) {
+szone_ptr_in_use_enumerator(task_t task, void *context, unsigned type_mask, vm_address_t zone_address, memory_reader_t reader, vm_range_recorder_t recorder)
+{

     szone_t            *szone;
     kern_return_t      err;
     szone_t            *szone;
     kern_return_t      err;

+    

     if (!reader) reader = _szone_default_reader;
     if (!reader) reader = _szone_default_reader;

-//    malloc_printf("Enumerator for zone %p\n", zone_address);

     err = reader(task, zone_address, sizeof(szone_t), (void **)&szone);
     if (err) return err;
     err = reader(task, zone_address, sizeof(szone_t), (void **)&szone);
     if (err) return err;

-//    malloc_printf("Tiny ptrs enumeration for zone %p\n", zone_address);
-    err = tiny_in_use_enumerator(task, context, type_mask,
-      (vm_address_t)szone->tiny_regions, szone->num_tiny_regions, szone->tiny_bytes_free_at_end , reader, recorder);
+    err = tiny_in_use_enumerator(task, context, type_mask, szone, reader, recorder);

     if (err) return err;
     if (err) return err;

-//    malloc_printf("Small ptrs enumeration for zone %p\n", zone_address);
-    err = small_in_use_enumerator(task, context, type_mask,
-      (vm_address_t)szone->small_regions, szone->num_small_regions, szone->small_bytes_free_at_end , reader, recorder);
+    err = small_in_use_enumerator(task, context, type_mask, szone, reader, recorder);

     if (err) return err;
     if (err) return err;

-//    malloc_printf("Large ptrs enumeration for zone %p\n", zone_address);

     err = large_in_use_enumerator(task, context, type_mask,
       (vm_address_t)szone->large_entries, szone->num_large_entries, reader,
       recorder);
     if (err) return err;
     err = large_in_use_enumerator(task, context, type_mask,
       (vm_address_t)szone->large_entries, szone->num_large_entries, reader,
       recorder);
     if (err) return err;

-//    malloc_printf("Huge ptrs enumeration for zone %p\n", zone_address);

     err = huge_in_use_enumerator(task, context, type_mask,
       (vm_address_t)szone->huge_entries, szone->num_huge_entries, reader,
       recorder);
     err = huge_in_use_enumerator(task, context, type_mask,
       (vm_address_t)szone->huge_entries, szone->num_huge_entries, reader,
       recorder);


@@ -2913,9 +3864,11 @@ szone_ptr_in_use_enumerator(task_t task, void *context,
 
 // Following method is deprecated:  use scalable_zone_statistics instead
 void
 
 // Following method is deprecated:  use scalable_zone_statistics instead
 void

-scalable_zone_info(malloc_zone_t *zone, unsigned *info_to_fill, unsigned count) {
+scalable_zone_info(malloc_zone_t *zone, unsigned *info_to_fill, unsigned count)
+{

     szone_t    *szone = (void *)zone;
     unsigned   info[13];
     szone_t    *szone = (void *)zone;
     unsigned   info[13];

+    

     // We do not lock to facilitate debug
     info[4] = szone->num_tiny_objects;
     info[5] = szone->num_bytes_in_tiny_objects;
     // We do not lock to facilitate debug
     info[4] = szone->num_tiny_objects;
     info[5] = szone->num_bytes_in_tiny_objects;


@@ -2934,54 +3887,71 @@ scalable_zone_info(malloc_zone_t *zone, unsigned *info_to_fill, unsigned count)
 }
 
 static void
 }
 
 static void

-szone_print(szone_t *szone, boolean_t verbose) {
-    unsigned   info[13];
-    unsigned   index = 0;
-    SZONE_LOCK(szone);
-    scalable_zone_info((void *)szone, info, 13);
-    malloc_printf("Scalable zone %p: inUse=%d(%y) touched=%y allocated=%y flags=%d\n", szone, info[0], info[1], info[2], info[3], info[12]);
-    malloc_printf("\ttiny=%d(%y) small=%d(%y) large=%d(%y) huge=%d(%y)\n", info[4], info[5], info[6], info[7], info[8], info[9], info[10], info[11]);
-    // tiny
-    malloc_printf("%d tiny regions: \n", szone->num_tiny_regions);
-    while (index < szone->num_tiny_regions) {
-       tiny_region_t   *region = szone->tiny_regions + index;
-       print_tiny_region(verbose, *region, (index == szone->num_tiny_regions - 1) ? szone->tiny_bytes_free_at_end : 0);
-       index++;
-    }
-    if (verbose) print_tiny_free_list(szone);
-    // small
-    malloc_printf("%d small regions: \n", szone->num_small_regions);
-    index = 0; 
-    while (index < szone->num_small_regions) {
-       small_region_t  *region = szone->small_regions + index;
-       print_small_region(szone, verbose, region, (index == szone->num_small_regions - 1) ? szone->small_bytes_free_at_end : 0);
-       index++;
-    }
-    if (verbose) print_small_free_list(szone);
-    SZONE_UNLOCK(szone);
+szone_print(szone_t *szone, boolean_t verbose)
+{
+  unsigned  info[13];
+  size_t    index;
+  region_t  region;
+  
+  SZONE_LOCK(szone);
+  scalable_zone_info((void *)szone, info, 13);
+  _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX,
+                 "Scalable zone %p: inUse=%d(%y) touched=%y allocated=%y flags=%d\n",
+                 szone, info[0], info[1], info[2], info[3], info[12]);
+  _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX,
+                 "\ttiny=%d(%y) small=%d(%y) large=%d(%y) huge=%d(%y)\n",
+                 info[4], info[5], info[6], info[7], info[8], info[9], info[10], info[11]);
+  // tiny
+  _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX,
+                 "%d tiny regions:\n", szone->num_tiny_regions);
+  for (index = 0; index < szone->num_tiny_regions_allocated; ++index) {
+    region = szone->tiny_regions[index];
+    if (region)
+      print_tiny_region(verbose, region, (region == szone->last_tiny_region) ? 
+                                         szone->tiny_bytes_free_at_end : 0);
+  }
+  if (verbose)
+    print_tiny_free_list(szone);
+  // small
+  _malloc_printf(MALLOC_PRINTF_NOLOG | MALLOC_PRINTF_NOPREFIX,
+                 "%d small regions:\n", szone->num_small_regions);
+  for (index = 0; index < szone->num_small_regions_allocated; ++index) {
+    region = szone->small_regions[index];
+    if (region)
+      print_small_region(szone, verbose, region,
+                         (region == szone->last_small_region) ? 
+                         szone->small_bytes_free_at_end : 0);
+  }
+  if (verbose)
+    print_small_free_list(szone);
+  SZONE_UNLOCK(szone);

 }
 
 static void
 }
 
 static void

-szone_log(malloc_zone_t *zone, void *log_address) {
-    szone_t    *szone = (void *)zone;
+szone_log(malloc_zone_t *zone, void *log_address)
+{
+    szone_t    *szone = (szone_t *)zone;
+

     szone->log_address = log_address;
 }
 
 static void
     szone->log_address = log_address;
 }
 
 static void

-szone_force_lock(szone_t *szone) {
-//    malloc_printf("szone_force_lock\n");
+szone_force_lock(szone_t *szone)
+{

     SZONE_LOCK(szone);
 }
 
 static void
     SZONE_LOCK(szone);
 }
 
 static void

-szone_force_unlock(szone_t *szone) {
-//    malloc_printf("szone_force_unlock\n");
+szone_force_unlock(szone_t *szone)
+{

     SZONE_UNLOCK(szone);
 }
 
 boolean_t
     SZONE_UNLOCK(szone);
 }
 
 boolean_t

-scalable_zone_statistics(malloc_zone_t *zone, malloc_statistics_t *stats, unsigned subzone) {
-    szone_t *szone = (void *)zone;
+scalable_zone_statistics(malloc_zone_t *zone, malloc_statistics_t *stats, unsigned subzone)
+{
+    szone_t *szone = (szone_t *)zone;
+    

     switch (subzone) {
        case 0: 
            stats->blocks_in_use = szone->num_tiny_objects;
     switch (subzone) {
        case 0: 
            stats->blocks_in_use = szone->num_tiny_objects;


@@ -3010,11 +3980,22 @@ scalable_zone_statistics(malloc_zone_t *zone, malloc_statistics_t *stats, unsign
 }
 
 static void
 }
 
 static void

-szone_statistics(szone_t *szone, malloc_statistics_t *stats) {
-    stats->blocks_in_use = szone->num_tiny_objects + szone->num_small_objects + szone->num_large_objects_in_use + szone->num_huge_entries;
-    size_t     big_and_huge = szone->num_bytes_in_large_objects + szone->num_bytes_in_huge_objects;
+szone_statistics(szone_t *szone, malloc_statistics_t *stats)
+{
+    size_t     big_and_huge;
+    
+    stats->blocks_in_use =
+      szone->num_tiny_objects +
+      szone->num_small_objects +
+      szone->num_large_objects_in_use +
+      szone->num_huge_entries;
+    big_and_huge = szone->num_bytes_in_large_objects + szone->num_bytes_in_huge_objects;

     stats->size_in_use = szone->num_bytes_in_tiny_objects + szone->num_bytes_in_small_objects + big_and_huge;
     stats->size_in_use = szone->num_bytes_in_tiny_objects + szone->num_bytes_in_small_objects + big_and_huge;

-    stats->max_size_in_use = stats->size_allocated = szone->num_tiny_regions * TINY_REGION_SIZE + szone->num_small_regions * SMALL_REGION_SIZE + big_and_huge ; 
+    stats->max_size_in_use = stats->size_allocated =
+      szone->num_tiny_regions * TINY_REGION_SIZE +
+      szone->num_small_regions * SMALL_REGION_SIZE +
+      big_and_huge ; 
+

     // Now we account for the untouched areas
     stats->max_size_in_use -= szone->tiny_bytes_free_at_end;
     stats->max_size_in_use -= szone->small_bytes_free_at_end;
     // Now we account for the untouched areas
     stats->max_size_in_use -= szone->tiny_bytes_free_at_end;
     stats->max_size_in_use -= szone->small_bytes_free_at_end;


@@ -3032,36 +4013,33 @@ static const struct malloc_introspection_t szone_introspect = {
 }; // marked as const to spare the DATA section
 
 malloc_zone_t *
 }; // marked as const to spare the DATA section
 
 malloc_zone_t *

-create_scalable_zone(size_t initial_size, unsigned debug_flags) {
+create_scalable_zone(size_t initial_size, unsigned debug_flags)
+{

     szone_t            *szone;
     szone_t            *szone;

-    vm_address_t       addr;
-    size_t             msize;
-    size_t             msize_used = 0;
-    // malloc_printf("=== create_scalable_zone(%d,%d) - %s\n", initial_size, debug_flags, (DEBUG_MALLOC) ? "**** DEBUG" : "");
-#if PAGE_SIZE_FIXED
-    if ((1 << vm_page_shift) == vm_page_size) {
-       // malloc_printf("vm_page_shift validated to be %d\n", vm_page_shift);
-    } else {
-       malloc_printf("*** vm_page_shift incorrectly set to %d\n", vm_page_shift);
+
+    /*
+     * Sanity-check our build-time assumptions about the size of a page.
+     * Since we have sized various things assuming the default page size,
+     * attempting to determine it dynamically is not useful.
+     */
+    if ((vm_page_size != _vm_page_size) || (vm_page_shift != _vm_page_shift)) {
+       malloc_printf("*** FATAL ERROR - machine page size does not match our assumptions.\n");

        exit(-1);
     }
        exit(-1);
     }

-#else
-    if (!vm_page_shift) {
-       unsigned        page;
-       vm_page_shift = 12; // the minimal for page sizes
-       page = 1 << vm_page_shift;
-       while (page != vm_page_size) { page += page; vm_page_shift++;};
-    }
-#endif
-    addr = allocate_pages(NULL, SMALL_REGION_SIZE, SMALL_BLOCKS_ALIGN, 0, VM_MAKE_TAG(VM_MEMORY_MALLOC));
-    if (!addr) return NULL;
-    szone = (void *)addr;
-    msize = (sizeof(szone_t) + SMALL_QUANTUM - 1) >> SHIFT_SMALL_QUANTUM;
-    // malloc_printf("sizeof(szone_t)=%d msize for 1st block=%d; wasted %d bytes\n", sizeof(szone_t), msize, (msize << SHIFT_SMALL_QUANTUM) - sizeof(szone_t));
-    small_meta_header(szone)[0] = msize;
+
+    /* get memory for the zone, which is now separate from any region.
+       add guard pages to prevent walking from any other vm allocations
+       to here and overwriting the function pointers in basic_zone. */
+    szone = allocate_pages(NULL, SZONE_PAGED_SIZE, 0, 
+                           SCALABLE_MALLOC_ADD_GUARD_PAGES, 
+                           VM_MEMORY_MALLOC);
+    if (!szone)
+        return NULL;
+    /* set up the szone structure */

     szone->tiny_regions = szone->initial_tiny_regions;
     szone->small_regions = szone->initial_small_regions;
     szone->tiny_regions = szone->initial_tiny_regions;
     szone->small_regions = szone->initial_small_regions;

-    msize_used += msize; szone->num_small_objects++;
+    szone->num_tiny_regions_allocated = INITIAL_NUM_REGIONS;
+    szone->num_small_regions_allocated = INITIAL_NUM_REGIONS;

     szone->basic_zone.version = 3;
     szone->basic_zone.size = (void *)szone_size;
     szone->basic_zone.malloc = (void *)szone_malloc;
     szone->basic_zone.version = 3;
     szone->basic_zone.size = (void *)szone_size;
     szone->basic_zone.malloc = (void *)szone_malloc;


@@ -3073,7 +4051,9 @@ create_scalable_zone(size_t initial_size, unsigned debug_flags) {
     szone->basic_zone.batch_malloc = (void *)szone_batch_malloc;
     szone->basic_zone.batch_free = (void *)szone_batch_free;
     szone->basic_zone.introspect = (struct malloc_introspection_t *)&szone_introspect;
     szone->basic_zone.batch_malloc = (void *)szone_batch_malloc;
     szone->basic_zone.batch_free = (void *)szone_batch_free;
     szone->basic_zone.introspect = (struct malloc_introspection_t *)&szone_introspect;

+    szone->debug_flags = debug_flags;

     LOCK_INIT(szone->lock);
     LOCK_INIT(szone->lock);

+

 #if 0
 #warning CHECK_REGIONS enabled
     debug_flags |= CHECK_REGIONS;
 #if 0
 #warning CHECK_REGIONS enabled
     debug_flags |= CHECK_REGIONS;


@@ -3082,16 +4062,7 @@ create_scalable_zone(size_t initial_size, unsigned debug_flags) {
 #warning LOG enabled
     szone->log_address = ~0;
 #endif
 #warning LOG enabled
     szone->log_address = ~0;
 #endif

-    szone->debug_flags = debug_flags;
-    szone->small_regions[0] = addr >> SMALL_BLOCKS_ALIGN;
-    szone->num_small_regions = 1;
-    msize_t    free_msize = NUM_SMALL_BLOCKS - msize;
-    small_meta_header(szone)[msize] = free_msize;
-    szone->small_bytes_free_at_end = free_msize << SHIFT_SMALL_QUANTUM;

     CHECK(szone, __PRETTY_FUNCTION__);
     CHECK(szone, __PRETTY_FUNCTION__);

-#if 0
-    write(1, "Malloc szone created\n", 23);
-#endif

     return (malloc_zone_t *)szone;
 }
 
     return (malloc_zone_t *)szone;
 }
 


@@ -3108,12 +4079,14 @@ create_scalable_zone(size_t initial_size, unsigned debug_flags) {
  *    malloc does not store flags in front of large page-aligned allocations.
  * 3) Original szone-based freezedrying code.
  * 4) Fresher malloc with tiny zone
  *    malloc does not store flags in front of large page-aligned allocations.
  * 3) Original szone-based freezedrying code.
  * 4) Fresher malloc with tiny zone

+ * 5) 32/64bit compatible malloc
+ * 6) Metadata within 1MB and 8MB region for tiny and small

  *
  * No version backward compatibility is provided, but the version number does
  * make it possible for malloc_jumpstart() to return an error if the application
  * was freezedried with an older version of malloc.
  */
  *
  * No version backward compatibility is provided, but the version number does
  * make it possible for malloc_jumpstart() to return an error if the application
  * was freezedried with an older version of malloc.
  */

-#define MALLOC_FREEZEDRY_VERSION 4
+#define MALLOC_FREEZEDRY_VERSION 6

 
 typedef struct {
     unsigned version;
 
 typedef struct {
     unsigned version;


@@ -3122,24 +4095,29 @@ typedef struct {
 } malloc_frozen;
 
 static void *
 } malloc_frozen;
 
 static void *

-frozen_malloc(szone_t *zone, size_t new_size) {
+frozen_malloc(szone_t *zone, size_t new_size)
+{

     return malloc(new_size);
 }
 
 static void *
     return malloc(new_size);
 }
 
 static void *

-frozen_calloc(szone_t *zone, size_t num_items, size_t size) {
+frozen_calloc(szone_t *zone, size_t num_items, size_t size)
+{

     return calloc(num_items, size);
 }
 
 static void *
     return calloc(num_items, size);
 }
 
 static void *

-frozen_valloc(szone_t *zone, size_t new_size) {
+frozen_valloc(szone_t *zone, size_t new_size)
+{

     return valloc(new_size);
 }
 
 static void *
     return valloc(new_size);
 }
 
 static void *

-frozen_realloc(szone_t *zone, void *ptr, size_t new_size) {
+frozen_realloc(szone_t *zone, void *ptr, size_t new_size)
+{

     size_t     old_size = szone_size(zone, ptr);
     void       *new_ptr;
     size_t     old_size = szone_size(zone, ptr);
     void       *new_ptr;

+    

     if (new_size <= old_size) {
        return ptr;
     }
     if (new_size <= old_size) {
        return ptr;
     }


@@ -3151,11 +4129,13 @@ frozen_realloc(szone_t *zone, void *ptr, size_t new_size) {
 }
 
 static void
 }
 
 static void

-frozen_free(szone_t *zone, void *ptr) {
+frozen_free(szone_t *zone, void *ptr)
+{

 }
 
 static void
 }
 
 static void

-frozen_destroy(szone_t *zone) {
+frozen_destroy(szone_t *zone)
+{

 }
 
 /********* Pseudo-private API for emacs unexec ************/
 }
 
 /********* Pseudo-private API for emacs unexec ************/


@@ -3173,8 +4153,9 @@ frozen_destroy(szone_t *zone) {
  * returns 0 (error) if any non-szone zones are encountered.
  */
 
  * returns 0 (error) if any non-szone zones are encountered.
  */
 

-int
-malloc_freezedry(void) {
+uintptr_t
+malloc_freezedry(void)
+{

     extern unsigned malloc_num_zones;
     extern malloc_zone_t **malloc_zones;
     malloc_frozen *data;
     extern unsigned malloc_num_zones;
     extern malloc_zone_t **malloc_zones;
     malloc_frozen *data;


@@ -3190,9 +4171,11 @@ malloc_freezedry(void) {
     data->nszones = malloc_num_zones;
     data->szones = (szone_t *) calloc(malloc_num_zones, sizeof(szone_t));
 
     data->nszones = malloc_num_zones;
     data->szones = (szone_t *) calloc(malloc_num_zones, sizeof(szone_t));
 

-    /* Fill in the array of szone structures.  They are copied rather than
+    /*
+     * Fill in the array of szone structures.  They are copied rather than

      * referenced, since the originals are likely to be clobbered during malloc
      * referenced, since the originals are likely to be clobbered during malloc

-     * initialization. */
+     * initialization.
+     */

     for (i = 0; i < malloc_num_zones; i++) {
        if (strcmp(malloc_zones[i]->zone_name, "DefaultMallocZone")) {
            /* Unknown zone type. */
     for (i = 0; i < malloc_num_zones; i++) {
        if (strcmp(malloc_zones[i]->zone_name, "DefaultMallocZone")) {
            /* Unknown zone type. */


@@ -3203,12 +4186,13 @@ malloc_freezedry(void) {
        memcpy(&data->szones[i], malloc_zones[i], sizeof(szone_t));
     }
 
        memcpy(&data->szones[i], malloc_zones[i], sizeof(szone_t));
     }
 

-    return (int) data;
+    return((uintptr_t)data);

 }
 
 int
 }
 
 int

-malloc_jumpstart(int cookie) {
-    malloc_frozen *data = (malloc_frozen *) cookie;
+malloc_jumpstart(uintptr_t cookie)
+{
+    malloc_frozen *data = (malloc_frozen *)cookie;

     unsigned i;
 
     if (data->version != MALLOC_FREEZEDRY_VERSION) {
     unsigned i;
 
     if (data->version != MALLOC_FREEZEDRY_VERSION) {


@@ -3235,3 +4219,4 @@ malloc_jumpstart(int cookie) {
 
     return 0;
 }
 
     return 0;
 }

+